If You’re Going to Scare People About AI, Help Them Through It
Most of my talks about AI are to CEOs and management teams and usually in person. Recently, I spoke to a group of students and faculty over Zoom. Afterward, I was told how anxious the students were about their future. They hadn’t raised specific concerns during the talk. I wish I’d known beforehand. It made me rethink what students need from a talk like this.
In some of the classes I’m taking at Massachusetts Institute of Technology, the mood seems almost the opposite. Those classes focus on recursive self-improvement, essentially AI helping improve AI. That’s a particular group, so I wouldn’t generalize from it. Still, the contrast made me think harder about who is listening when we talk about this technology.
A CEO may see an opportunity to get more done. Someone looking for a first job may wonder how they’ll get experience if AI does the work they would have learned from. Telling that person to be adaptable doesn’t answer the question.
The job market is already difficult. The New York Fed’s data show unemployment among recent college graduates rising from 4.7 percent in June 2024 to 5.7 percent in June 2026. About 42 percent of employed recent graduates were working in jobs that typically don’t require a college degree. That underemployment rate was already 41.2 percent in June 2019. These figures don’t tell us how much, if any, of the change is due to AI. [7]
We keep telling people they’ll need judgment to check AI’s work. Well, where does that judgment come from? You develop it by doing the work, making decisions and having someone explain what you missed. MIT professor Sasha Rakhlin also cautions against assuming that judgment or asking good questions will always be beyond AI’s reach. [6]
Some junior work is busywork. Some is how you learn to spot a bad assumption or a conclusion the evidence doesn’t support. If a junior researcher uses AI to help gather papers, I still want them to explain which studies belong in the review and why. Have them compare their choices with the AI’s and discuss the disagreements with a supervisor. Keep a record of why an approach was rejected. We need to know what the person learned, along with whether the review got finished.
That’s a problem worth working on now, regardless of whose jobs forecast turns out to be right.
Anthropic CEO Dario Amodei predicted in 2025 that AI could displace half of entry-level white-collar jobs within one to five years. In his January 2026 essay, he also discussed possible responses. Knowing how to build a model doesn’t tell us how quickly employers will change their businesses or what new work they’ll create. [1]
I want lab leaders to be honest about risk. But if you’re going to scare the shit out of everyone, help them through it. You’re building this technology. Helping people adjust, and helping pay for it, should be part of running the company.
The media have a responsibility here, too. Axios used “white-collar bloodbath” in a headline, although the article also discussed responses. [2] I’m tired of coverage built around shock. Ask where the numbers come from, explain the assumptions and go back later to check. A CEO can understand the technology and still have a commercial interest in how the story gets told.
To be fair – There are efforts to help. Anthropic has committed $200 million to external research on economic responses, including how apprenticeships and mentoring could develop expertise if AI absorbs junior tasks. [3] OpenAI is funding 14 independent policy projects, with results expected in 2027. [4] I’m glad they’re funding this work. It will take time to know what works, but we should be able to see what they’re trying and who it reaches.
If you’re hiring someone, think about what they’ll learn while using AI. Give that person real decisions to make, someone to review them and time to understand the corrections. Give experienced employees time to teach. A graduate shouldn’t need an unpaid year or another expensive degree to get started.
Educators have a related problem. In a randomized study at one Turkish high school, students using a basic GPT-based assistant did better on practice problems but worse on a subsequent exam without it than students who hadn’t used AI. A tutor designed with teacher input largely removed that harm, though it didn’t improve unaided exam performance over the control group. [5]
Schools should check what students can explain and solve independently, as well as what they produce with AI. Handing someone a chatbot isn’t an education strategy.
AI companies should help fund paid placements with employers and universities. Have someone independent check what participants can do, whether they get and keep jobs, and whether they’re taking on more demanding work. Report what fails, too. The labs should also tell us what AI is doing to junior hiring and training in their own companies.
I’m optimistic about what people can do with this technology. But telling someone that society will eventually adjust is little comfort when they need a job next year. Adjustment takes work, money and decisions about who gets an opportunity.
I can’t promise students their chosen profession will be safe. I can help them understand the technology and keep building their knowledge. The people building and deploying AI have a much bigger responsibility. If they believe these changes are coming as quickly as they say, helping people through them should be part of running the company.
If you can’t lead that work yourself, hire someone who can. Someone willing to challenge you, not another yes person. Give them the money and authority to do the work, and show us whether it’s helping. But please stop just yelling fire. Help people figure out what to do next.
Sources and context
[1] Dario Amodei, The Adolescence of Technology, January 2026
https://darioamodei.com/essay/the-adolescence-of-technology
Amodei’s jobs forecast and proposed responses. The forecast is a prediction, not an observed employment result.
[2] Axios, May 28, 2025
https://www.axios.com/2025/05/28/ai-jobs-white-collar-unemployment-anthropic
[3] Anthropic Economic Futures Research Fund agenda, July 22, 2026
https://www.anthropic.com/news/economic-futures-research-fund-agenda
Company funding commitment and research agenda, including apprenticeships and mentoring. These are proposed areas of research, not demonstrated transition outcomes.
[4] OpenAI, New policy ideas for the Intelligence Age, August 17, 2026
https://openai.com/index/new-policy-ideas-for-the-intelligence-age/
Company announcement of fourteen independent policy projects, with results expected in 2027.
[5] Bastani et al., Generative AI without guardrails can harm learning, PNAS, 2025
https://www.pnas.org/doi/10.1073/pnas.2422633122
Nearly 1,000 students at one Turkish high school. Basic GPT assistance improved practice but harmed later unaided exam performance. The teacher-informed tutor largely removed that harm without an exam gain over the control group.
[6] Stefanie Koperniak, interview with Sasha Rakhlin, MIT News, October 8, 2026
https://news.mit.edu/2026/3-questions-best-path-forward-for-ai-academia-1008
Rakhlin’s analysis of graduate training, individual contributions and changing AI capabilities.
[7] Federal Reserve Bank of New York, The Labor Market for Recent College Graduates, data updated August 6, 2026
https://www.newyorkfed.org/research/college-labor-market
Uploaded data workbook, observations through June 2026. Unemployment: unemployed!D425 and D449. Underemployment: underemployed!B365 and B449. These series do not isolate AI’s effect.


