Most Companies Have an AI Policy. Few Can Stop a Model

AI governance theatre is governance that produces the evidence of control without the power to change what an AI system does. You can test for it with one question: has this governance ever stopped, changed or delayed a deployment, and could it do so this week?

Most organisations would struggle to say yes. In ISACA's 2026 AI Pulse Poll, released on 5 May 2026, over 3,400 people working in IT audit, governance, cybersecurity and privacy described their own organisations. Sixty-eight percent said their employer had an AI policy, whether formal or limited. Twelve percent said there was a documented process for shutting down or overriding an AI system that is also tested regularly. An untested process is closer to a hope, so I read that 12 percent as the share that could show it can stop a system.

My position will annoy some governance teams. I would rather inherit a two-page AI policy with a shutdown drill that has actually been run than a forty-page policy with no drill. The document says what should happen. The drill shows that someone can make it happen, and how long it takes.

How wide is the gap between AI policy and control?

The ISACA figures are reported by professionals about their own employers, so they carry the usual error of any self-report. They still show a clear shape. Policy has become common, and the ability to act on a failing system has not caught up.

Other surveys point the same way. In McKinsey's survey of 1,363 people in early 2024, 18 percent said their organisation had an enterprise-wide council or board with the authority to make decisions on responsible AI governance. In its next survey, run in July 2024 with 1,491 participants, 17 percent said their board of directors oversaw AI governance. A policy needs someone with authority behind it before it can change anything, and most of these organisations had not given that authority to an enterprise-wide body or to the board.

Where does the idea of governance theatre come from?

The security writer Bruce Schneier calls the original version security theater: "security primarily designed to make you feel more secure." His favourite setting is the airport. In 2015 auditors from the Department of Homeland Security's inspector general ran covert tests of passenger screening at eight airports. The inspector general, John Roth, told Congress that November that the results were classified, and that they were "disappointing and troubling." He added: "We found layers of security simply missing." ABC News, citing officials briefed on the results, reported that screeners failed 67 of 70 tests. The inspector general has never published a figure.

Schneier has also argued that theater can do real work when it calms a fear out of proportion to the risk. AI governance has the opposite problem. The risk is real, and the reassurance arrives before the controls do.

The passengers in those queues did everything the process asked of them. The process looked complete, and it missed the thing it existed to catch. AI governance can fail the same way while the people running it are entirely sincere.

Theatre is a different problem from the limits every framework has, and it is not a new idea: the ethics-washing debate and Michael Power's older work on audit as a ritual of verification describe the same habit. What the test below adds is a question a board can ask on a Tuesday. Even honest governance cannot see everything, as I set out in I Built Five Governance Frameworks. Here Is What They Cannot Do. Theatre is governance that has stopped trying to act at all.

What are the five kinds of AI governance theatre?

I group governance theatre into five kinds. Each produces an artefact that looks like governance, and each fails the same test in its own way.

Ethics board theatre. On 26 March 2019 Google announced an eight-member Advanced Technology External Advisory Council to "consider some of Google's most complex challenges that arise under our AI Principles." It was to meet four times that year and publish a report summarising its discussions. The announcement gave it no power to decide anything. On 4 April Google added a note to the post: "It's become clear that in the current environment, ATEAC can't function as we wanted. So we're ending the council and going back to the drawing board." An advisory body can be useful. If it has never changed a launch and has no route to stop one, it is commentary.

Policy theatre. A policy that lives only in a document depends on people reading it and choosing to follow it. It becomes a control when a system enforces it: an approved model gateway, a deployment check that fails when a required review is missing, a register that keeps an unregistered system out of production. In ISACA's poll a quarter of organisations had no active policy at all, so the policy is often the first thing a programme builds. The mistake is stopping there. I make the longer case for putting governance inside the tools people already use in The Best AI Governance Is Invisible.

Assessment theatre. An impact assessment can be filed without anything waiting on it. The EU AI Act shows the pattern in law. Article 27 requires certain deployers of high-risk systems, including public bodies and private companies providing public services, to assess and report the impact on fundamental rights before first use. Once the assessment is done, the deployer "shall notify the market surveillance authority of its results." The article contains no approval step, and nothing in it makes the deployment wait for an answer. Other parts of the Act do gate deployment; this one only reports. An assessment can still change a design if the people writing it are allowed to say no. Ask how many of yours ever did. Good upstream review shapes proposals before they arrive, so a high pass rate can be healthy. If every assessment in two years passed unchanged, with no conditions attached and no scope narrowed, ask whether the process is recording decisions rather than making them.

Disclosure theatre. A study of 32,111 model cards on Hugging Face, published in February 2024, found the training section filled in on 74.3 percent of cards, but the limitations section on only 17.4 percent and evaluation on 15.4 percent. The parts a user needs to judge risk were the parts most often left blank. Disclaimers fail the same way. In March 2024 The Markup found that New York City's MyCity chatbot, built on Microsoft's Azure AI services, told business owners they could take a cut of their workers' tips and told landlords they did not need to accept tenants with housing vouchers, both against city law. Its page warned that it "may occasionally produce incorrect, harmful or biased content." The city added a warning not to use its answers as legal or professional advice and kept the bot online. It came down in February 2026, after Mayor Zohran Mamdani called it "functionally unusable." A warning label moves the risk to the user without telling them which answers are wrong.

Consultation theatre. A feedback route only counts if what comes through it can change the outcome. Australia's Robodebt scheme raised welfare debts by averaging annual tax-office income across fortnights. Approving the class-action settlement in 2021, the Federal Court of Australia found that the Commonwealth had unlawfully asserted debts of at least AU$1.763 billion against about 433,000 people. It also found that the Commonwealth "maintained the Robodebt system in the face of approximately 29 decisions of the Administrative Appeals Tribunal (AAT) up to 30 May 2017" that rejected the method. The appeal route worked one case at a time, and the scheme ran until the government conceded a test case in November 2019. I tell that story, with the case for writing stop rules before launch, in Nobody Tells You When to Stop an AI Deployment.

What does governance that can stop a model look like?

Substance is rarer, but you can see it when it happens. In June 2022 Microsoft said it would retire facial analysis capabilities in its Azure Face service "that purport to infer emotional states and identity attributes such as gender, age, smile, facial hair, hair, and makeup." New customers lost access that day, and existing customers had until 30 June 2023. Microsoft tied the decision to its updated Responsible AI Standard and gave its reasons, among them "the lack of consensus on a definition of 'emotions'." It kept the features for accessibility uses such as its Seeing AI app, and the account of why comes from Microsoft itself. Still, governance removed a working capability from a product customers were already using, which is the kind of evidence the test looks for.

The law is moving the same way. From 2 December 2027, high-risk systems listed in Annex III of the EU AI Act, which covers uses such as hiring and credit scoring, must reach deployers in a form that lets the people overseeing them, "as appropriate and proportionate", "intervene in the operation of the high-risk AI system or interrupt the system through a 'stop' button or a similar procedure that allows the system to come to a halt in a safe state" (Article 14(4)(e)). The date comes from the Digital Omnibus, in force since 27 July 2026. Breaches of the high-risk obligations of providers and deployers carry fines of up to EUR 15 million or 3 percent of worldwide annual turnover under Article 99(4). The Regulation Tracker keeps these dates current.

Five things separate a governance programme that can stop a model from one that performs:

• A named stop authority. One person who can pause a named system without the business owner's permission, written down before launch, and protected from being overruled or sidelined for using it.
• Triggers agreed in advance. The error rate, complaint count, overturned appeals or drift that pauses the system without a debate on the day.
• A rehearsed stop. A drill with a measured time from decision to halt, and a known fallback for the work the system was doing.
• A tested switch for bought systems. For software you buy, the stop is a contract right and a feature switch you have tried, because you do not control the vendor's model.
• A record of changes. Launches delayed, scopes narrowed, systems retired. An empty record after a year is itself the finding.

Agents make this harder, because what an agent has already done stays done when you switch it off. I cover that case in Your Board Can't Turn It Off.

Why could governance theatre become a legal risk?

Regulators have started testing what companies say about AI against what they do. On 18 March 2024 the SEC settled charges against two investment advisers, Delphia and Global Predictions, for $400,000 in total, over false and misleading statements about their use of AI. "Such AI washing hurts investors," said the SEC chair, Gary Gensler. Six months later the FTC announced Operation AI Comply, five actions against deceptive AI claims, including a $193,000 settlement with DoNotPay. "There is no AI exemption from the laws on the books," said the FTC chair, Lina Khan.

Both actions date from 2024, and enforcement priorities can shift when an agency's leadership changes. The EU fines above are written into law. The US cases were also about claims of AI capability. My reading, which no regulator has tested yet, is that published claims about how carefully a company governs its AI are statements of the same kind. A responsible AI page that promises human oversight is a representation. If nobody can stop the system, that page may become evidence against the company that published it.

Five questions for the next board meeting

A board can find out in one meeting whether its AI governance is substance or performance. These five questions do it.

1. When did our AI governance last stop, change or delay a deployment, and what was it?
2. Who can pause each of our highest-risk AI systems today, without asking the business owner?
3. When did we last rehearse a shutdown, and how long did it take from decision to halt?
4. Which rules in our AI policy are enforced by a system rather than by people remembering them?
5. What do we not know about the systems we run, and who is responsible for finding out?

If the answers arrive as a dashboard with every light green, ask again. A Board AI Dashboard That Is Always Green Is Broken sets out what a board should see instead. The fifth question is the hardest to answer honestly, and I explain why in I Do Not Fully Understand the AI I Help Govern.

An organisation starting from little can begin with minimum viable governance: an inventory, a risk register and a named owner. Add the stop authority to that list on the first day.

Two in three organisations now have an AI policy. The question that matters comes after it: when a model goes wrong on a Tuesday afternoon, who stops it, and how long does that take?

Sources

• ISACA, "AI Use Accelerates, While Governance and ROI Lag, Says New ISACA Research", press release on the 2026 AI Pulse Poll, 5 May 2026 (more than 3,400 digital trust professionals).
• Tamim Ahmed, "The AI Security Gap: Adoption Is Accelerating but Response Capability Is Lagging", ISACA Now blog, 5 May 2026.
• McKinsey, "The state of AI in early 2024", 30 May 2024 (1,363 participants, 22 February to 5 March 2024).
• Alex Singla, Alexander Sukharevsky, Lareina Yee and others, "The state of AI: How organizations are rewiring to capture value", McKinsey, March 2025 (1,491 participants, 16 to 31 July 2024).
• Bruce Schneier, "In Praise of Security Theater", Wired, 25 January 2007.
• Michael Power, The Audit Society: Rituals of Verification, Oxford University Press, 1997.
• John Roth, Inspector General, Department of Homeland Security, statement to the House Committee on Oversight and Government Reform, 3 November 2015.
• Mike Levine and Justin Fishel, "DHS Internal Watchdog 'Deeply Concerned' Over TSA's Screening at Airports", ABC News, 9 June 2015.
• Kent Walker, "An external advisory council to help advance the responsible development of AI", Google, 26 March 2019, updated 4 April 2019.
• Regulation (EU) 2024/1689 (the AI Act), Articles 14, 27 and 99, as amended by Regulation (EU) 2026/1744, Official Journal, 24 July 2026.
• Weixin Liang, Nazneen Rajani and others, "What's documented in AI? Systematic Analysis of 32K AI Model Cards", arXiv, 7 February 2024.
• Colin Lecher, "NYC's AI Chatbot Tells Businesses to Break the Law", The Markup, 29 March 2024; Colin Lecher, Katie Honan and Maria Puertas, follow-up, 2 April 2024; Colin Lecher and Katie Honan, "Mamdani to Kill the NYC AI Chatbot", 30 January 2026.
• Federal Court of Australia, Prygodicz v Commonwealth of Australia (No 2) [2021] FCA 634, 11 June 2021.
• Sarah Bird, "Responsible AI investments and safeguards for facial recognition", Microsoft Azure blog, 21 June 2022.
• US Securities and Exchange Commission, press release 2024-36, 18 March 2024.
• US Federal Trade Commission, "FTC Announces Crackdown on Deceptive AI Claims and Schemes", 25 September 2024.

All sources checked on 6 October 2026.

Changes on 6 October 2026. Rewritten from about 5,700 words as a Reference piece with a new headline. The opening statistics now come from ISACA's 2026 poll. The old dek's "95 percent" TSA figure is now attributed to ABC News, because the inspector general classified the results and never published a number. Removed: survey figures that could not be traced to the cited source (a 36 percent framework figure credited to Pacific AI; the 18 percent council figure is McKinsey's, and is now credited to it), vendor and consultancy blog statistics, the Anthropic, Clearview, Gemini and hiring-bias cases (cut for length), five hand-built charts, and a download box pointing at a file that did not exist.