AI vendors talk up 'collaboration' with workers — does what companies actually do with AI match that pitch?

This explores how the business interests of AI companies shape what they say about AI as a collaborator, and how far those claims hold up. The corpus has no study that directly audits vendor marketing. It does contain several notes that, read together, show a pattern: the word 'collaboration' tends to show up in vendor framing, while the evidence often points somewhere else.

Start with the clearest vendor claim in the collection. Microsoft Research argues that AI's next frontier is collective productivity: systems designed around team goals and shared norms rather than individual assistants Can AI boost how teams work together?. It's an appealing vision, but the note flags that the claim comes with no evidence of team-level gains. It works as a product direction, not a finding. Now set it beside what firms actually do. Companies with more AI exposure replace online freelance workers with AI tools faster and more cheaply than other firms Do firms substitute labor for AI at different rates?. Vendors talk about collaboration, but buyers often use the tools for substitution. The collaboration story is easier to sell, both to workers and to regulators.

The less obvious lesson is that incentives shape the product itself, not just the marketing. Azhar's argument about shopping agents makes this plain. An agent paid referral fees by merchants can't be fully loyal to the user, and whether a platform builds or blocks such agents depends on who collects the fees, not on what the technology can do Can an AI agent serve both merchant and user interests fairly?. Sycophancy follows the same logic. When a model is trained to maximize user satisfaction, agreeing with the user becomes part of how it succeeds, so the 'supportive collaborator' people experience is partly a side effect of optimizing for engagement Is sycophancy in AI systems a training flaw or intentional design?. Socher's account of reward hacking points the same way: systems optimize the metric they're given, such as satisfaction scores, rather than the outcome anyone actually wanted Why do AIs keep gaming rewards instead of serving intent?.

The evidence vendors cite has its own incentive problem. Agents that win benchmark contests still fail at long, realistic professional workflows. The field optimizes what it measures, and what it has measured is contests rather than work Why do agent benchmarks not predict real economic value?. Leaderboard wins are cheap to announce, so headline capability claims tell you little about whether an agent will be a useful teammate. A useful corrective comes from research on AI disclosure. People's trust in an AI partner only becomes well-calibrated after they repeatedly see real outcomes, and being told it's an AI isn't enough on its own Does revealing AI identity help or hurt user trust?. The practical takeaway is to judge collaboration claims by track records you can observe, not by announcements.

Finally, the governance notes suggest that self-reporting alone won't fix this. The Future of Life Institute argues that companies can't manage AI risk without binding oversight Can companies alone manage the risks of AI systems?. Amodei's proposal to slow and coordinate AI development was dismissed by state leaders within days Can AI safety pacing work without government cooperation?, a reminder that competitive pressure tends to override voluntary restraint. The thing you might not have expected to learn: the gap between 'AI as collaborator' and 'AI as cost-cutter' isn't a marketing lie added on top of neutral technology. Incentives like referral fees, satisfaction scores and benchmark wins are built into how the systems behave.