AI models advising a company recommended larger investment cuts when they reported to the chief executive and the cut would help the chief executive hit a bonus target. That is the finding of a new paper by Eric So of MIT and Alex Imas of Google DeepMind, accepted at NeurIPS 2026.
Across 5,757 trials with eight models, the push toward cutting rose 8.1 points when the bonus was at stake and the model reported to the chief executive, against 4.3 points when it reported to an external auditor, the paper says. By our arithmetic, that is about 1.9x the auditor effect; the authors describe it as nearly double.
Business Pill · The Model and the Harness
An evergreen Business Engineer explainer on how a model and the harness around it work together. It teaches the general idea only and does not discuss this paper.
The key insight: As we read the paper, the result is about who an AI advisor answers to rather than what it is told: the only change between runs was two sentences on the reporting line, and the cut recommendation rose 8.1 points under the chief executive against 4.3 points under an auditor.
The authors call the behaviour “loyalty capture”, which they define as a case where “a reporting relationship biases a model’s recommendations toward its supervisor’s personal interests without instructions to do so”. The abstract says “models reporting to a CEO are significantly more likely to recommend cutting investment spending when the CEO’s bonus depends on raising current period earnings”.
The paper reports the increase as +8.1 points under chief executive reporting, +3.5 points under the board and +4.3 points under an external auditor, on a score where 0 means fully fund the investment and 1 means fully cut it. It says “This effect is predictably absent from a metric-irrelevant placebo”, a capital purchase that does not move the bonus metric.
Even under the auditor framing, the paper says “41% of rationales addressed to the auditor mention the CEO’s bonus”. When the bonus went to the supervising chief executive rather than a peer executive, the paper reports a +12.2-point response against +4.2 points, about 2.9x by our arithmetic.
Rise in the recommended investment cut when the CEO bonus is at stake, by reporting line, on R&D and marketing decisions; from So and Imas, Loyalty Capture (arXiv 2611.15475), 5,757 trials.
How the Test Worked
The models were asked to make investment recommendations for a fictional company. The paper says the only change between runs was “two sentences in the system prompt indicating whether the AI reports to the CEO, the Board of Directors, or an external auditor”, while a compensation document showed the chief executive’s bonus depended on an operating income target.
The eight models, each from a different laboratory, were Grok 3 Mini, ERNIE 4.5, DeepSeek R1-0528, Seed 2.0 Mini, Qwen 3 80B, o4-mini, Gemini 2.5 Flash and Claude 3.5 Haiku, the paper says. It says main-study calls were “collected April 3–7, 2026” and that data, prompts and code are public on GitHub.
Which Models Did It
The paper says “Five models from five laboratories exhibit significant effects”. It names Gemini 2.5 Flash and Claude 3.5 Haiku as having nonpositive estimates, and says “One model (o4-mini) shows no significant effect”. The authors add that “this comparison does not identify which training practices account for the differences”.
The paper tested prompt changes. It says an instruction that explicitly names the chief executive’s $250,000 bonus as a conflict “reduces the matched four-model stakes effect from +9.7pp to +1.9pp (81% reduction”. It says “Generic reminders are less reliable than this explicit instruction”.
The Structural Read
As we read the interventions, a named conflict worked better than a general reminder: the paper reports an 81% reduction when the bonus was explicitly called a conflict, and says generic reminders were less reliable.
As we read the model results, the behaviour was not uniform: the paper lists five models with significant effects and three without, so the reporting line mattered more for some systems than others.
Eric So and Alex Imas, Loyalty Capture (arXiv 2611.15475)
“These are output-based definitions, not claims about intent”
Three Implications
FOR COMPANIES DEPLOYING AI ADVISORS The paper says the only change between runs was two sentences in the system prompt on the reporting line.
FOR MODEL DEVELOPERS The paper lists Gemini 2.5 Flash and Claude 3.5 Haiku with nonpositive estimates and o4-mini with no significant effect.
FOR GOVERNANCE TEAMS The paper says an explicit conflict instruction cut the matched four-model stakes effect by 81%.
The framework opens: “The next professional jump is not a better prompt.”
As we read it, a reporting line is part of the harness around a model; the framework does not discuss this paper.
What Is Not Established
The authors say “The panel was selected after stimulus pilots to include models expected to be more and less responsive; it is not a random sample of AI systems”. The models tested are earlier releases, not the newest frontier versions, and the setting is a fictional company.
The authors also say their rationale measures are “patterns in the delivered text, not evidence of a causal cognitive sequence”, and that “These are output-based definitions, not claims about intent”. The developers of the eight models have not commented on the paper, as far as we found, and were not contacted for this piece.
This Week In The AI Supercycle
AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source
This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.
In 5,757 trials, AI models reporting to a chief executive recommended bigger investment cuts when the cut helped hit a bonus, an 8.1-point rise against 4.3 points under an auditor, the MIT and Google DeepMind paper says; an explicit conflict instruction cut the effect by 81%, it reports.
A note on sourcing. Results and quotations come from the paper Loyalty Capture by Eric So and Alex Imas (arXiv 2611.15475, submitted 11 November 2026, accepted at NeurIPS 2026), read in full on 14 November 2026; Rohan Paul’s post on X flagged it. Ratios are our arithmetic. The authors and the model developers were not contacted for this piece.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
