The Burden of Proof
Making intelligence cheap has a strange side effect.
You get more answers, more content, more predictions, more automated decisions. More companies calling themselves AI companies, more products labelled intelligent, more policies claiming readiness, more systems telling us what they know. And then a different
problem shows up.
How do we know any of it deserves to be believed?
That question kept following me through this issue. Not every article is about truth in the obvious sense; we have venture capital, semiconductors, employment, cybersecurity, legal responsibility, AI regulation and search. But underneath them sits something harder to ignore. As AI makes production easier, the burden shifts toward proving what is real, what works, what came from where, and who is prepared to stand behind it.
Trust is the feeling. Proof is what justifies it. This issue is mostly about the second.
Prof. Anton I. Botha, in this month’s featured piece, finds the gap first inside organisations. Companies can have AI policies while their employees are already several steps beyond them, which is why his case for HR leadership has less to do with who owns the software than with whether the organisation understands what people are actually doing with it. The policy says one thing. Behaviour often says another, and behaviour is usually the more useful evidence.
Evane Alexandre comes at the same ground through Article 50 of the EU AI Act. Certain
AI-generated content will increasingly carry machine-readable markings designed to make its synthetic origin detectable. Straightforward enough, until you ask where that information leads. Depending on implementation, provenance may become linkable to a service, an account, a generation event, potentially an individual. A transparency mechanism can also become an attribution mechanism. Whether synthetic content can be detected is the easier half of the question. What the infrastructure remembers once it can is the harder one.
Proof is never neutral. The moment something can be traced, someone has to decide what is retained, who can resolve the link, and when it can be used.
Luana Lo Piccolo pulls that thread into responsibility. When increasingly autonomous systems take part in decisions, actions and workflows, tracing an outcome back to meaningful human agency gets harder. We still want a name at the end of the chain: someone who exercised control, understood the consequence, and answers when something goes wrong. AI is making those apparently simple questions less simple.
Wesley Lin makes it almost brutally practical. The Cyber Resilience Act can require a vulnerability report within 24 hours. A supplier can have policies, documentation and assurances, and none of that proves it can detect the problem, route the information to the right person and file before the clock runs out. His supplier-readiness test is really a test of whether governance survives contact with reality. A procedure on paper and a procedure that holds under pressure are two different animals.
Shalini Gopalkrishnan widens the frame. If some forms of cognitive output become extraordinarily cheap, output itself becomes weaker proof of value. An answer, a draft, an analysis may cost almost nothing to produce. Deciding whether it is reliable, appropriate and worth acting on does not. The cheaper generation gets, the more valuable judgment becomes.
Then the argument moves down the stack. Lucy Jenyi Chang looks at semiconductor companies whose histories and market identities were set in one era, while AI infrastructure changes which capabilities count. What looked peripheral can become strategic once power, manufacturing, packaging or efficiency turns into the constraint. Strip away the label and the capability underneath tells you more.
Milad Khademi does something similar with capital. “AI investment” has become so large a category that the label risks telling us very little, so his data looks underneath the headline and asks where the money is actually moving. Applications made AI visible. Increasingly, the capital story also runs through models, compute and infrastructure. Watch what investors choose to fund underneath the headline number.
By the end of the issue, misinformation feels almost too narrow a word for what is happening here. The real problem is evidence.
A polished answer is cheap. So is a convincing document, a published policy, an AI label, an attached citation, a claim of readiness. When appearances become cheap, proof becomes expensive.
That may be the harder problem now. The scarce resource may not be intelligence itself, but our ability to establish why something should be trusted, where it came from, whether it works, and who will take responsibility for it.
The burden of proof has not disappeared. AI may simply be moving it somewhere else.
My thanks to every contributor, reviewer, editor and volunteer who made this September issue possible. And a special thank you to Prof. Anton I. Botha for joining us as this month’s Featured contributor and for bringing a distinctly human and organisational perspective to the question of AI adoption.
