buildOne report1 publisher China's rules for human-like AI, in force since July 15, 2026, require companion chatbots to end a conversation promptly when a user asks to stop. Teams shipping companion products there now have to rebuild the goodbye turn and decide how distress gets escalated.
Reality
- Evidence50
- Adoption
- Insufficient
- Hype gap+5
- Incentives
- Insufficient
- Confidence55
Six authors writing for the ACM's Technology Policy Council say AI has made code and vulnerability reports cheap to produce while the people who decide what ships are volunteers at projects like Apache, whose 2024 revenue was $2,379,402.
Reality
- Evidence55
- Adoption20
- Hype gap+10
- Incentives60
- Confidence55
An Entrepreneur contributor column asks managers to list the ten recurring tasks behind a requisition and sort them into judgment, relationship, repetition and coordination work. Its supporting studies measured support agents and consultants.
Reality
- Evidence33
- Adoption28
- Hype gap+22
- Incentives38
- Confidence55
Harvard and Alberta researchers coded 54,861 subreddit posts around two real AI updates and surveyed 1,452 people. The reaction to Replika's silent change was larger, and it lasted longer than the reaction to GPT-5.
Reality
- Evidence58
- Adoption50
- Hype gap+12
- Incentives60
- Confidence55
An NBER working paper asked 916 doctoral students and postdocs how they felt now and how they felt half a year earlier. The gap is large, and it measures stated intent, which is not the same as measuring departures.
Reality
- Evidence56
- Adoption
- Insufficient
- Hype gap+20
- Incentives68
- Confidence50
Gates says technology executives are privately worried about AI's labor effects and are telling each other to stay quiet to protect fundraising, which leaves the people who must staff 2027 planning without a policy timetable.
Reality
- Evidence34
- Adoption31
- Hype gap+16
- Incentives74
- Confidence42
At roughly $87 a week, the humans show up once; anything a learner wants to repeat is answered by a render. That split, not the model behind it, is the product.
Reality
- Evidence34
- Adoption28
- Hype gap+14
- Incentives62
- Confidence42
In a Harvard, MIT and University of Washington experiment, 228 reviewers deferred to an LLM at the same rate with or without a rationale. Only the silent version improved their accuracy.
Reality
- Evidence42
- Adoption
- Insufficient
- Hype gap+18
- Incentives34
- Confidence45