AI backlash reached a new pitch on 16 August 2026, when Anthropic chief executive Dario Amodei published a written response arguing that the anger aimed at his industry is not a communications failure at all. His diagnosis was one sentence long: “I think it is fundamentally a crisis of trust.” The remark was picked up by TechCrunch and Fortune the same day and has been argued over ever since, because it shifts the blame from how AI is described to what AI companies have actually delivered.
That distinction matters far beyond San Francisco. If you are being asked to approve an AI strategy this year, the AI backlash is now part of your operating environment: it shapes what your staff will accept, what your customers will forgive, and what your regulators will ask for. Amodei’s argument is that the scepticism is structural and decades old, not something a better press release can fix.
This article sets out exactly what he wrote, the investor argument that provoked it, the survey data that supports the trust reading of the AI backlash, and — most usefully — what a mid-sized business should change in its own AI rollout as a result. The evidence is stronger than the headlines suggest, and the practical lessons are not the ones the industry keeps repeating.
Table of contents
- What Anthropic’s CEO Actually Said About the AI Backlash
- Why This AI Backlash Is a Crisis of Trust, Not a Messaging Problem
- The Investor Argument Behind the AI Backlash Row
- What the 2026 Polling Says About the AI Backlash
- The AI Backlash Is Not Only an American Story
- The AI Backlash Has a Physical Address: Data Centres
- Regulation, Open Weights and the False Choice
- The Undelivered Promise Is the Real AI Backlash Lesson
- What the AI Backlash Means for Your Own AI Projects
- A Trust-First Checklist for Any AI Rollout
- How to Read Vendor Promises After the AI Backlash
- Frequently Asked Questions About the AI Backlash
- References and Further Reading
What Anthropic's CEO Actually Said About the AI Backlash
The response was posted on X and written as a short two-part essay rather than a thread of one-liners, which is Amodei’s habit when he engages publicly. Three claims carry the argument.
The core diagnosis
Amodei’s central line on the AI backlash is that distrust of AI is downstream of distrust in institutions generally. “I think that ordinary people don’t trust companies, governments, or the tech industry and always suspect that we are cooking up some new way to screw them over,” he wrote. Nothing in that sentence is specific to machine intelligence. It describes a public that has been primed by decades of broken promises to read every new technology as a scheme.
The rejection of the messaging theory
He also disputed the premise that his own warnings created the problem. His public writing has been “about equally balanced between risks and benefits,” he argued, pointing to Machines of Loving Grace — his 2024 essay on AI’s upside — which he says he wrote because he did not feel the industry was painting an inspiring enough picture of how the technology could transform the world for the better. On that reading, the AI backlash cannot be pinned on doom-laden marketing by a single chief executive.
The admission
The most quotable part is the concession. “By far the most accurate criticism of AI companies including Anthropic is that we haven’t yet delivered on our big promises to benefit the world,” Amodei wrote, adding that this is “totally on us, and I think it’s the criticism you should be making, instead of all this stuff about messaging and marketing.” He was equally direct about what would fix it: “The thing that will work is actually curing cancer.” Repeating the promise, he suggested, has become a cliché rather than an inspiration.
The honesty argument
Underneath all three claims sits a position on disclosure that is worth reading twice, because it is the part with direct commercial application. “Honesty is the right thing on the merits, and in terms of public credibility and trust it is no worse than, and may in fact be better than, an approach that ignores or distracts from risks.” That is a testable claim about buyer behaviour, not a philosophical one.
Why This AI Backlash Is a Crisis of Trust, Not a Messaging Problem
The distinction between a trust problem and a messaging problem sounds academic until you notice how differently each one is solved. A messaging problem is fixed by changing what you say. A trust problem is only fixed by changing what you do and then waiting.
Trust is earned on the last promise, not the next one
Every AI vendor is currently asking the public to believe a forward-looking claim. The AI backlash exists because the public is still marking the previous one. Social platforms promised connection and delivered engagement optimisation. Cloud providers promised simplicity and delivered egress bills. Each unredeemed promise raises the evidence threshold for the next pitch, which is why an identical claim from an identical company lands worse in 2026 than it did in 2023.
Distrust is cheap and correct often enough
Scepticism persists because it is frequently vindicated. When a technology’s benefits are speculative and its costs are immediate — energy bills, job insecurity, data exposure — assuming the worst is a reasonable default rather than an irrational one. That asymmetry is the engine of the AI backlash, and it is not addressed by louder optimism.
Which means the fix is delivery, not narrative
If Amodei is right, the only durable response to the AI backlash is a stack of small, verifiable, boring wins. That is unglamorous advice for a frontier lab and extremely practical advice for a business deploying AI internally. A rollout that saves a team four hours a week and can prove it does more for trust than any all-hands presentation about transformation.
The Investor Argument Behind the AI Backlash Row
The exchange did not appear from nowhere. It was a reply to Gavin Baker, an investor who had made the opposite case on the All-In podcast and on X.
What Baker argued
Baker’s position was that Amodei’s persistent warnings about catastrophic AI risk have themselves fuelled the AI backlash — particularly the wave of local opposition to data centre construction in the United States — and that Amodei had effectively already lost the regulatory argument. The prescription that followed was that the head of a company likely to become enormously significant should spend his public voice defending the industry rather than cataloguing its dangers. Baker also repeated a secondhand claim that Amodei had privately mused Anthropic could become “the only private company in the world.”
Where the argument is strongest
Baker is not obviously wrong that rhetoric has consequences. Risk language from credible insiders does travel further than risk language from critics, and it does get recycled by campaigners with unrelated objectives. If you are running a public-affairs strategy, treating your own chief executive’s warnings as free ammunition for your opponents is a rational worry.
Where it falls down
The difficulty is causal ordering. The trends that constitute the AI backlash — falling institutional trust, job-loss anxiety, resistance to industrial development near homes — were all measurable before any frontier lab existed, and they move in countries where Amodei is unknown. Attributing them to one executive’s essays requires the public to be paying far more attention to AI-safety discourse than any survey suggests it does.
| Baker’s claim | Amodei’s answer | What the evidence favours |
|---|---|---|
| Risk warnings caused the public turn against AI | Distrust predates AI and targets institutions generally | Amodei — the trend lines start before 2023 |
| His messaging is skewed toward danger | Roughly equal on risks and benefits; cites his 2024 upside essay | Mixed — the risk material is quoted far more often |
| The regulatory fight is already lost | Workable rules can cover risk and concentration together | Unresolved — depends on the jurisdiction |
| Regulation concentrates power in big labs | Calls that a false choice; backs revenue thresholds and open weights | Amodei — thresholds already exist in California’s SB 53 |
| He should defend the industry instead | Honesty is no worse for credibility than distraction | Amodei — but the claim is still largely untested |
What the 2026 Polling Says About the AI Backlash
Amodei’s trust reading is checkable, and the survey record broadly supports it. Two large 2026 studies of US opinion point the same way.
The Pew numbers
Pew Research Center surveyed 5,119 US adults between 17 and 23 February 2026. Sixty-three per cent said AI is advancing too quickly, against 2% who said too slowly. Seventy-one per cent said AI makes their personal information less secure, against 3% who said more secure. Sixty-seven per cent had little or no confidence in the US government to regulate AI effectively, up from 62% in 2024, and around 59% had no confidence in US companies to develop and use it responsibly. Asked about the next twenty years, 40% expected a negative effect on society against 16% expecting a positive one.
The adoption paradox
The same study found 49% of US adults now use AI chatbots, up from 33% in 2024. That combination — rising use alongside falling confidence — is the most important single fact about the AI backlash. People are not boycotting the technology. They are using it while expecting to be exploited by it, which is a far harder position to sell into than simple refusal.
The Gallup direction of travel
A Bentley University–Gallup survey of 3,270 US adults, run from 4 to 11 May 2026 with a margin of error of ±2.4 points, shows the AI backlash getting worse year on year rather than settling. Thirty-nine per cent now say AI does more harm than good, up from 31% in 2025. Seventy-nine per cent expect AI to reduce the number of US jobs over the next decade, up from 73%. And the share who trust businesses to use AI responsibly fell from 31% to 27% — the first time that measure has moved in the wrong direction since the series began.
The generational split nobody planned for
Among 18-to-29-year-olds the numbers are worse on every measure: 47% say AI does more harm than good, up from 36% a year earlier; 75% expect job losses, up thirteen points; and trust in business use fell from 30% to 20%. The cohort with the highest AI fluency is the most hostile to it, which quietly demolishes the industry’s favourite explanation that the AI backlash is a familiarity problem that education will solve.
The AI Backlash Is Not Only an American Story
British readers sometimes treat the AI backlash as an American argument about American data centres. UK survey evidence says otherwise, with one important difference in what the public asks for.
What UK respondents want
A nationally representative survey of 3,513 UK residents by the Ada Lovelace Institute and the Alan Turing Institute found 72% saying that laws and regulation would increase their comfort with AI, up from 62% in 2023. Eighty-nine per cent backed an independent regulator with real enforcement powers. Around 74% had used a chatbot by spring 2026, so this is not a population speaking from inexperience.
Why that shape of demand matters commercially
The British version of the AI backlash is less “stop building this” and more “show me who is accountable when it goes wrong.” That is a request for governance, and governance is something a supplier can actually provide. A named owner, a documented review process and a route to challenge an automated decision are cheaper than a marketing campaign and considerably more persuasive.
The AI Backlash Has a Physical Address: Data Centres
The clearest place to watch the AI backlash convert from opinion into cost is planning permission. Sentiment is hard to price; a stalled development is not.
The Q1 2026 numbers
Data Center Watch counted more than 75 US data centre projects blocked or delayed in the first quarter of 2026, with an estimated combined value of around $130 billion — roughly equal to everything blocked or delayed across the whole of 2025. Organised opposition grew from 396 groups at the end of 2025 to 833 active in 49 states by March. More than 300 relevant bills were introduced in state legislatures in the first six weeks of the year.
What the objections are actually about
The complaints are rarely about superintelligence. They concern electricity prices, water consumption, noise, land use and the observation that a very large building employs very few people. This is the AI backlash in its least abstract form, and it is the strongest evidence for Amodei’s argument: the objection is to a concrete imposition with no visible local return, exactly the pattern that produces institutional distrust in any sector.
The read-across for anyone running infrastructure
If you operate or buy data centre capacity, the practical consequence of the AI backlash is that siting and expansion timelines now carry political risk that did not need budgeting for two years ago. Contracts signed on the assumption of frictionless build-out are the ones most exposed.
| Measure | 2025 | Q1 2026 |
|---|---|---|
| Value of US projects blocked or delayed | ~$130bn across twelve months | ~$130bn in three months |
| Active opposition groups | 396 at year end | 833 by March |
| States with organised opposition | Not reported at this level | 49 |
| Legislative activity | Mostly incentive-led | 300+ bills in six weeks, oversight-led |
| Dominant objection | Energy cost | Energy, water, noise, low local employment |
Regulation, Open Weights and the False Choice
Baker’s sharpest point was that safety regulation entrenches incumbents. Amodei’s reply is the most policy-dense part of the exchange, and it is where a business reader can see which way compliance is heading.
The three-part test he proposes
He called the choice between restricting frontier development and distributing capability widely a “false choice,” arguing that the right “rules of the road” can simultaneously “(a) address AI’s cyber/bio/alignment risks, (b) institutionally constrain the power of the frontier AI companies, and (c) leave room for open-weights models while also addressing the specific risks that they bring.” He also restated a view that cuts against his own commercial interest: AI is a technology that structurally tends to concentrate power.
The mechanisms he named
Three are worth noting because they are already live proposals rather than abstractions. He praised California’s SB 53 for exempting companies below a revenue threshold, so obligations scale with size. He endorsed pre-deployment testing. And he floated a FINRA-style self-regulatory body — an industry-funded supervisor with statutory backing — as an alternative to either a new federal agency or nothing at all.
What to take from it operationally
Revenue thresholds and pre-deployment testing are the two ideas most likely to reach you. If thresholds become the norm, your obligations depend on your size rather than your risk appetite, and documentation becomes the cheapest form of insurance. Pre-deployment testing, meanwhile, is simply an evaluation gate — something every serious internal AI project should already have.
The Undelivered Promise Is the Real AI Backlash Lesson
Strip away the personalities and one sentence survives as the useful part of the whole exchange: the criticism that lands is that the promises have not yet been kept.
Why “curing cancer” is a bad pitch
Amodei’s own example makes the point against the genre. A promise so large that no one can verify progress against it functions as noise. Every year it goes unredeemed, it deepens the AI backlash rather than easing it, because the audience learns that scale of claim carries no information. The same dynamic ruins internal AI programmes pitched as company-wide transformation.
Verifiable smallness beats unverifiable bigness
The alternative is claims sized so that failure would be obvious. Reducing average first-response time on support tickets from nine hours to three is checkable. Cutting manual invoice matching by 60% is checkable. Neither will make a keynote, and both build the kind of credibility that survives the AI backlash because a sceptic can audit them.
The honesty test
Amodei’s claim that honesty about risk is “no worse than, and may in fact be better than” avoidance is the most directly transferable idea here. Publishing your AI system’s known failure modes to the staff using it feels risky and reliably produces better outcomes: people calibrate, they catch errors, and they stop treating the tool as either infallible or worthless. Our own trust and security practice starts from the same premise.
What the AI Backlash Means for Your Own AI Projects
Most organisations reading about the AI backlash conclude that it is someone else’s fight. It is not. The scepticism arrives inside your business through three doors.
Your staff
Seventy-nine per cent of US adults expect AI to cut jobs, and your employees are part of that number. Any rollout presented without a clear answer on job security will be quietly resisted, and quiet resistance is expensive: shadow workarounds, unreported errors, and adoption figures that look fine in the dashboard while nothing changes in the work. Deciding whether AI agents augment or replace a role — and saying so — is not an HR nicety. It is the adoption plan.
Your customers
With only 27% of people trusting businesses to use AI responsibly, disclosure is now a competitive variable. Telling a customer that an AI system drafted a reply, and that a human checked it, costs one sentence. Being discovered not to have said it costs considerably more, and the AI backlash makes that discovery likelier because people are actively looking.
Your regulators and buyers
The UK demand for accountability shows up in procurement long before it shows up in law. Enterprise security questionnaires already ask which models process which data, where inference happens and who reviews automated decisions. Firms that cannot answer lose deals for reasons that never get recorded as an AI backlash problem.
Your cost base
There is a fourth door that is easy to miss: spend. Distrust makes people demand oversight, oversight adds review steps, and review steps consume the savings the project was justified on. Building AI cost governance in from the start is what stops a trust-driven control layer from turning a positive business case negative.
A Trust-First Checklist for Any AI Rollout
None of the following requires a new platform. All of it is the operational translation of Amodei’s argument that delivery and honesty, not narrative, are what shift the AI backlash.
Name a human owner for every automated decision
Not a team, a person. The Ada Lovelace finding that 89% want an enforceable accountability mechanism is a demand for a name. Publish it internally and give staff a route to challenge an output without going through the project’s sponsor.
Publish the failure modes
Write down what the system gets wrong, how often, and what to do when it happens. A support assistant that reads a customer’s message relies on natural language processing to infer intent, and that inference is wrong often enough to need a documented escape hatch. Saying so up front is the single cheapest trust intervention available.
Measure something a sceptic would accept
Pick one metric that existed before the project, that nobody controls, and that would visibly fail if the tool did not work. Report it monthly whether it flatters the project or not.
Keep a human in the loop where the cost of error is asymmetric
Automate the reversible, review the irreversible. Refunds, dismissals, credit decisions and clinical or legal outputs belong in the review lane regardless of measured accuracy, because the AI backlash is driven by consequence, not by error rate.
Disclose to customers in plain words
One sentence, in the channel where the AI was used, in language a non-technical reader understands. Avoid “AI-powered.” Say what happened and who checked it.
Write down the data path
Which model, hosted where, trained on what, retained how long. Pew’s finding that 71% believe AI makes their personal information less secure means this is the specific fear you are answering, and a vague answer confirms it.
| Trust signal | What it looks like in practice | Typical effort |
|---|---|---|
| Named accountable owner | One person per automated decision, published internally | Hours |
| Documented failure modes | Known error types, rates and the escalation route | Days |
| Independent success metric | A pre-existing number nobody on the project owns | Days |
| Customer-facing disclosure | One plain sentence at the point of use | Hours |
| Documented data path | Model, hosting region, training basis, retention period | Weeks |
| Human review on irreversible actions | A defined review lane with capacity budgeted | Ongoing cost |
How to Read Vendor Promises After the AI Backlash
The most useful side-effect of the AI backlash is that it gives buyers permission to ask harder questions. Amodei’s own framing supplies the test: does this supplier make claims that would be obviously falsified if untrue?
The four questions worth asking
Ask what the system gets wrong and how often, and treat “it’s very accurate” as a non-answer. Ask which specific number in your business will move, and by when. Ask what happens to your data, naming the hosting region. And ask for a reference customer of your size in your sector, not a logo wall. Comparing options across the AI models and tools landscape is easier once you insist on falsifiable answers.
The red flags
A vendor that will not describe failure modes has either not measured them or does not want you to. A vendor whose business case rests entirely on headcount reduction is handing you a staff-relations problem alongside the software. And a vendor that cannot name where inference happens cannot help you answer the procurement questionnaire you will face.
| Vendor promise | Question that tests it | Red flag answer |
|---|---|---|
| “Highly accurate” | Accurate on what task, measured how, against what baseline? | A benchmark score with no link to your workflow |
| “Enterprise-grade security” | Which model, hosted in which region, retaining data how long? | “It’s all encrypted” |
| “Transformational ROI” | Which existing metric moves, by how much, by when? | A percentage with no named metric behind it |
| “Replaces a whole team” | What is the review process for its mistakes? | No review lane described at all |
| “Used by leading brands” | Can we speak to a customer our size in our sector? | A logo wall and no reference call |
Frequently Asked Questions About the AI Backlash
Did Amodei blame the public for the AI backlash?
No. His argument runs the other way: the distrust is a reasonable response to institutions that have not delivered, and the fix is on the companies. He named the failure to deliver on big promises as the most accurate criticism of AI firms including his own.
Is the AI backlash actually reducing AI use?
Not so far. Pew found chatbot use rising from 33% to 49% of US adults between 2024 and 2026 while confidence fell. The AI backlash is currently expressed as suspicion during use, plus organised resistance to visible infrastructure, rather than abstention.
Does the AI backlash affect small and mid-sized firms?
Yes, mostly through procurement and staff adoption rather than protest. Buyers ask harder AI questions, employees resist rollouts that dodge the job-security question, and both are cheaper to address before a project starts than after.
What is the FINRA-style body Amodei proposed?
An industry-funded self-regulatory organisation with statutory backing, modelled on the US financial-industry supervisor, as a middle path between a new federal agency and no oversight. It is a proposal, not a policy.
Is Anthropic’s position self-interested?
Partly, and Amodei concedes the shape of the objection by stating that AI structurally tends to concentrate power and by supporting revenue thresholds that would exempt smaller companies. Judge the specific mechanisms rather than the motive.
What should we do first?
Pick one AI use case already running, name its owner, write down its failure modes and disclose it to the affected staff and customers. That sequence answers the AI backlash where it actually reaches you, and it takes days rather than quarters. For a related view of how these promises are marketed, see our analysis of why people aren’t buying Mark Zuckerberg’s AI future.
References and Further Reading
TechCrunch: Anthropic CEO says AI backlash is ‘fundamentally a crisis of trust’
Fortune: Dario Amodei admits AI suffers from a crisis of trust
Dario Amodei: Machines of Loving Grace
Pew Research Center: Americans and AI in 2026
Pew Research Center: Key findings about how Americans view artificial intelligence
Gallup: Americans Cool Toward AI
Ada Lovelace Institute and Alan Turing Institute: How do people feel about AI?
Data Center Watch: Q1 2026 Report
NBC News: Data center opposition sharply rising in 2026
NIST: AI Risk Management Framework
ICO: Guidance on AI and data protection
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
