Compromised Witnesses

A developer tells a standup she's 40% more productive with Copilot. She's been saying this for three months. Her tech lead quotes the number in a capacity planning document. The document reaches a steering committee. The steering committee approves a six-figure enterprise licence. At no point does anyone measure whether the 40% shows up in the team's cycle time, defect rate, or cost per feature.

This is not a story about AI. It is a story about how organisations build conviction from compromised testimony.

Robert Cialdini identified commitment and consistency as one of the primary mechanisms of human persuasion. The principle is simple: once a person makes a public statement, they will work to remain consistent with it, even when the evidence shifts. The commitment doesn't have to be dramatic. It just has to be visible. Cialdini's research showed that written, public, freely chosen commitments are the most binding. A standup is all three.

Charlie Munger, who spent six decades thinking about cognitive error in the context of capital allocation, called this "commitment and consistency tendency" and placed it among the most dangerous biases in institutional decision-making. In Poor Charlie's Almanack, he observed that the brain conserves programming space by being reluctant to change. What begins as an offhand remark hardens into identity. The developer who said she's more productive becomes the developer who is more productive with AI. Reversing that position means reversing something about herself.

This isn't a character flaw. It's the normal operation of a well-documented cognitive mechanism. But it has specific consequences for how organisations evaluate AI tools, and those consequences are not widely understood.

The problem compounds in three ways.

The first is selection bias in reporting. Developers who find AI tools useful talk about it. Developers who tried them and stopped tend to say nothing, because publicly abandoning a tool the organisation is investing in carries social cost. The visible evidence skews positive. Not conspiracy. Just the predictable result of asymmetric incentives around disclosure.

The second is the absence of baseline measurement. Most engineering organisations cannot tell you their current cycle time per feature, their defect injection rate by author, or their total cost of rework. Without these numbers, the claim "AI makes us more productive" is unfalsifiable. It's not that the claim is false. There is simply no apparatus to test it. And unfalsifiable claims, as Karl Popper spent a career explaining, are not claims at all. They are articles of faith.

The third is what Cialdini would recognise as a cascade of social proof. Once enough people in an organisation have publicly committed to AI productivity, the uncommitted face pressure not from management but from the apparent consensus of their peers. Cialdini's research on social proof shows that people look to others' behaviour for guidance on what is correct, particularly under uncertainty. AI adoption in most engineering teams is nothing if not uncertain. The conditions for a social proof cascade are close to ideal.

Baldur Bjarnason, a software researcher whose The Intelligence Illusion deserves wider reading in engineering leadership, recently observed that LLM advocates are caught in Cialdini-style cognitive traps and have lost sight of the distinction between personal usefulness and organisational value. He identifies a taxonomy that most engineering leaders have not thought about carefully: personal usefulness, perceived productivity, project productivity, project ROI, and overall business value.

An AI tool can score well on the first two and be a disaster on the rest. A developer who generates code twice as fast but produces output that takes three times as long to review has improved her own experience at the expense of the team's throughput. She will not notice this. She will report herself as more productive, because from where she sits, she is. Her testimony is compromised not by dishonesty but by the limits of her vantage point, reinforced by the commitment she has already made.

Any quantitative researcher would recognise the problem. Subjective experience is a poor proxy for system-level effects. We know this. We have known it for decades. We simply keep forgetting it when the technology is new and exciting enough.

Lisanne Bainbridge's 1983 paper "Ironies of Automation" described a version of this problem in a different context. As automated systems handle more routine work, the human operators who remain become less practiced at the skills they need when the automation fails. The system degrades the competence of its own safety net.

The parallel to AI-assisted development is uncomfortable. If a developer delegates the generative work and retains the review work, she needs to be a better reviewer than before, not a worse one. But review is cognitively expensive and, as anyone who has sat through a code review will confirm, substantially less enjoyable than writing. The commitment bias compounds: having publicly said the tool makes her productive, she is now psychologically invested in approving its output efficiently. Slowing down to scrutinise contradicts the narrative she has committed to.

Daniel Kahneman's distinction between System 1 and System 2 thinking is relevant here. AI-generated code that looks plausible activates System 1: fast, pattern-matching, energy-conserving. Catching the subtle errors that distinguish working code from almost-working code requires System 2: slow, effortful, and exactly the kind of thinking that commitment bias discourages. The reviewer who has told herself she is more productive has a psychological incentive to stay in System 1.

None of this means AI tools are useless. It means that self-reported developer experience is a structurally unreliable input to an engineering leader's decision. The witnesses are compromised, not by malice but by the predictable operation of cognitive biases that Cialdini, Munger, and Kahneman mapped out long before anyone had heard of a large language model.

The engineering leader's job is to build measurement systems that do not depend on testimony. Cycle time. Defect density. Time to review. Rework rate. Cost per feature delivered. These are observable, and they do not have opinions about whether the tools are working. If the numbers improve after AI adoption, you have evidence. If they don't, you have an expensive subscription and a team that believes it is more productive than it is.

I have written before that trust, not capability, is the bottleneck in AI-assisted engineering. The Cialdini problem is the human half of that argument. You cannot trust the system if the only evidence you have is the system's users telling you it works, because the act of using it and publicly endorsing it compromises their ability to evaluate it. None of this is new in engineering or in science. It's the reason we have double-blind trials, independent audits, and separation of duties. We did not invent these disciplines because people are dishonest. We invented them because people are human.

The question for engineering leaders is straightforward: do you have measurement systems that would tell you if AI adoption were making your teams slower? If the answer is no, your current confidence is built on Cialdini, not data.

Sources and further reading

Robert Cialdini, Influence: The Psychology of Persuasion (1984, revised 2021). The commitment and consistency principle is Chapter 3. The social proof chapter is equally relevant to understanding how AI adoption spreads through organisations.

Charlie Munger, "The Psychology of Human Misjudgment" in Poor Charlie's Almanack (2005). Munger's treatment of commitment and consistency tendency in institutional settings is the best bridge between Cialdini's experimental psychology and the reality of corporate decision-making.

Daniel Kahneman, Thinking, Fast and Slow (2011). The System 1/System 2 framework explains why AI-generated output that looks plausible is harder to evaluate critically than output that looks obviously wrong.

Lisanne Bainbridge, "Ironies of Automation" (1983), Automatica, 19(6), 775-779. The original paper on how automation degrades the competence of the humans who oversee it. Still the most concise statement of the problem.

Baldur Bjarnason, The Intelligence Illusion (2023, second edition 2025). Bjarnason's taxonomy of usefulness levels (personal, perceived, project, ROI, business value) is the sharpest framework available for understanding why individual developer testimony diverges from organisational outcomes.

Karl Popper, The Logic of Scientific Discovery (1959). The falsifiability criterion. If you cannot specify what evidence would disprove the claim that AI is making your team more productive, you are not making an empirical claim.