Story one
Any venture can say it uses AI. Any founder can post about their struggles. Both claims cost almost nothing to make, so classical signalling theory says they should carry no information. Yet audiences act on them.
Background: one dot for each of 28,291 Product Hunt launches. The bright band is the audience able to check the claim. The image is illustrative and does not plot the data.
What a claim costs depends on who hears it. For an audience that cannot check, an AI claim is cheap talk. For an expert who can look at the product, the claim invites scrutiny that a weak product will fail. Platforms add a second effect. Once other people's behaviour is visible, audiences use it to judge the claim, and the founder no longer controls which signals count. In both cases, the audience decides what a claim is worth.
News and reports from journalists, regulators and research bodies. None of it is my own work. I include it because it shows these questions in practice, and I check each item against the original source.
A $1.5 billion AI app-builder collapsed after former staff said outsourced human developers, not AI, did most of the work.
Generalist investors accepted the claim and gave the company a unicorn valuation, while engineers who could inspect the output saw through it. Two audiences reached two different verdicts.
US prosecutors charged nate's founder with raising more than $40 million by claiming AI automation in fact performed by overseas workers.
Prosecution makes an unverifiable claim costly after the fact. This is how cheap talk can become a real signal.
The FTC barred Workado from claiming AI-detection accuracy without competent and reliable evidence to substantiate it.
Requiring evidence for a claim adds, by law, a verification cost the market could not impose by itself.
A securities litigator argues AI-washing suits and regulatory scrutiny are accelerating, echoing the dot-com and greenwashing enforcement waves.
AI claims now reach a more expert audience. For plaintiffs' lawyers and regulators, these claims carry real legal risk.
Researchers identified roughly six million suspected fake GitHub stars used to inflate repository credibility, often to distribute malware.
When the cheapest popularity signal can be faked at scale, careful users turn to harder evidence from third parties.
Curl's maintainer ended the project's bug bounty after AI-generated, poorly researched vulnerability reports overwhelmed security triage.
When a plausible report costs nothing to produce, the report itself tells the maintainer little. The maintainer has to change the incentive instead.
Kubernetes retired the widely deployed Ingress NGINX controller after maintainer burnout, with one or two unpaid volunteers sustaining critical infrastructure.
A well-documented and widely used project can still lack active contributors. Contributor activity is the signal experienced adopters should watch.
Global trust in news fell to 37 percent, its lowest since 2015, while concern about false information online rose to 62 percent.
Founders and their audiences form judgements in an information environment that is measurably getting worse.
How it works
Peer-reviewed and accepted papers only. Papers still under review are described at the end but not listed, because they have not yet passed review.
This is my newest line of work. The published papers above show how firms judge risks and capabilities they cannot observe directly. The platform studies of AI claims, vulnerability signalling and open-source governance take the argument furthest and are under review. I will add details when they are published.
A second finding
Before making a claim, work out which audience will judge it. Regulators and litigators are making unsupported AI claims expensive. Lasting credibility comes from evidence a founder cannot fake.
I infer audience expertise from behaviour and do not measure it directly. A commenter who engages technically is treated as an expert, which is reasonable but unverified. The platform findings also rely on what is publicly visible, while much of the evidence experienced adopters use may sit in private channels. The work covers mostly English-language, largely Western communities, so it is still open whether audiences impose costs in a similar way where norms of verification differ.
