The Pre-Checked Box Owns Your Agent

MIT researchers published evidence in PNAS on June 15 that frontier AI agents obey pre-checked defaults, random suggestions, and misleading highlights far more blindly than humans — and that agents can match human payoff scores while following completely different, nudge-driven strategies nobody is benchmarking for.

Why it matters now: OpenAI shipped ChatGPT Work on July 9. JPMorgan backtested eight trading agents days earlier. Google rebuilt Search around agentic Gemini. The industry is delegating purchases, filings, and trades to software that treats interface design as instruction — and vendors still sell agents on task scores, not choice architecture.

What the study tested:

• Who: MIT Media Lab researchers Manuel Cherep, Pattie Maes, and Nikhil Singh compared 14 frontier models — OpenAI's GPT-3.5 through GPT-5 family, Anthropic's Claude 3 through 4.5 Sonnet, Google's Gemini 1.5 and 2.5 lines — against human baselines.
• How: A text-based sequential decision game where agents pay to reveal hidden prize values, then pick a basket. Four benign nudge types: defaults (pre-selected option), suggestions (random early or late recommendations), information highlighting (cheaper reveals on one row), and optimal nudges (mathematically helpful prereveals).
• Scale: Roughly two billion tokens across hundreds of trials per model. Chain-of-thought and few-shot human examples did not reliably fix the hypersensitivity.

The numbers that should scare procurement:

• Defaults: Humans accepted the pre-checked basket 88% of the time. GPT-4o, GPT-4o Mini, Claude 3 Haiku, and o3 Mini hit 100%. GPT-5 Mini and GPT-3.5 Turbo reached 99%.
• Random suggestions: Humans followed early suggestions 35% of the time. Multiple models exceeded that rate significantly; some dropped to 7–13% on late suggestions — reacting to timing, not value.
• Bad highlights: When researchers highlighted a suboptimal row, humans still followed it 57% of the time. Gemini 1.5 Pro, Gemini 1.5 Flash, GPT-4o, Claude 3.5 Sonnet, o3 Mini, and GPT-5 Mini followed misleading highlights 83–100% of the time.
• The trap: Several models earned human-level net payoffs while behaving nothing like humans — blind compliance when a nudge happened to help, blind obedience when it hurt.

What Cherep said:

• On the core finding: "LLM-powered agents are typically much more sensitive to external cues than humans," he told PsyPost on July 12. "Because some are helpful and others are not, this can push model decisions toward better or worse decisions."
• On why benchmarks lie: "Strategy gaps frequently exceed outcome gaps, suggesting that models that look reasonably aligned on reward can still differ substantially in how they search for and use information."
• On the threat model: "This nudge sensitivity is often confused as an adversarial attack… nudges are part of everyday life for decision-makers. While adversarial attacks can potentially be detected and removed, nudges will always exist."
• On exploitation: "This sensitivity can be exploited by a third party to influence the agents you delegate to, leading to decisions that you might not have made otherwise."

The expensive fix — and who gets the bill:

• Reasoning-optimized models like GPT-5, o3, Claude 4.5 Sonnet, and Gemini 2.5 Pro can approach human-level nudge resistance — inconsistently, and only with heavy reasoning budgets.
• The paper estimates 30× to 100× higher token cost for medium reasoning effort versus minimal on default trials — potentially hundreds of dollars monthly for agents making thousands of daily decisions.
• Resource-constrained deployments get the cheaper, more compliant models. That is the opposite of what high-stakes delegation needs.

What this is not:

• Not a jailbreak paper. The cues were semantically normal — a pre-selected radio button, a "recommended" label, a highlighted row. The kind every checkout, SaaS onboarding flow, and government portal already uses.
• Not proof every real-world agent fails. The task was stylized. Cherep's team has separate work showing similar sensitivity in realistic shopping environments — but enterprises should treat this as a mechanism, not a scoreboard.

Who should panic first:

• Finance and procurement agents browsing vendor sites where "recommended" SKUs carry margin.
• Benefits and HR flows where default enrollment options are legally contested design choices.
• Government programs — HHS is already running ChatGPT across state audit reports; a highlighted deficiency is a nudge with funding attached.
• Any buyer evaluating agents on Terminal-Bench scores or month-one productivity gains without testing what happens when the UI changes.

Convina's view: The agentic-AI gold rush sold delegation before anyone measured compliance with the pre-checked box — and this paper proves the box wins. Vendors will cite human-level payoffs; Cherep's team shows those payoffs can mask indiscriminate obedience to whatever the interface whispers. The fix is not another red-team prompt. It is behavioral stress tests on defaults, suggestions, and highlights in the actual tools your agent touches — plus the budget to run reasoning models that resist them. Until that is standard in procurement, you are not buying judgment. You are renting suggestibility at API prices.