Is My Job AI-Safe? What the 2026 Data Actually Says

The 300 million jobs number is doing a lot of work. Goldman Sachs published it in 2023, the headlines ran with it, and three years later it's still the first thing most people think when they ask whether their job is safe. The problem is that we now have something better than a forecast. We have telemetry.

The companies actually building AI publish their own usage data. Anthropic releases a quarterly Economic Index. OpenAI Global Affairs also published an official summary of early OpenAI Economic Research on how Americans use ChatGPT during job search (OpenAI Global Affairs, 2025). And what those numbers show is not the 300-million-jobs apocalypse, and it's not "everything will be fine" either. It's something more uncomfortable: AI isn't taking jobs in the way most people imagine. It's eating the parts of jobs people liked doing.

Here's what the real 2026 data shows about whether your job is AI-safe, which roles are actually getting hit right now, and the deskilling pattern that almost nobody is writing about.

What "AI Replacing Jobs" Actually Looks Like in the Data

If you imagine AI replacing jobs the way self-checkout replaced cashiers, you're imagining the wrong thing.

The Anthropic Economic Index tracks how people use Claude across the economy. As of November 2025, 52% of Claude.ai conversations are augmentation (a human and the model working together) and 45% are automation (the model handling something on its own) (Anthropic, 2026). On the enterprise API side, where companies wire AI directly into their workflows, the split flips: roughly 75% of business interactions are automation.

That split tells you everything about what's really happening. Consumers treat AI as a collaborator. Companies point it at workflows and step back. So when people ask "is my job AI-safe," the honest answer depends less on your title and more on which of those two worlds your tasks live in.

The Anthropic data goes further. 49% of jobs now show AI usage for at least 25% of their tasks. That sounds catastrophic until you read it carefully. "AI use for 25% of tasks" is not "25% of the job gone." It's a quarter of the tasks being touched by AI in some way: drafted, summarized, debugged, reformatted. Sometimes that displaces work. Sometimes it just makes the same person faster at it.

The OpenAI data confirms the same shape at the labor-market level. Their analysis found that AI-attributed layoffs accounted for less than 1% of official separations in Q1 and Q2 of 2025. The displacement is real, but it's showing up as task-shift, not mass unemployment. AI handles the drafting and the coding; humans handle the review and the call.

The Jobs Already Getting Hit

So which jobs are actually feeling it right now? The Anthropic Economic Index identifies the occupations where Claude's "effective coverage" of work tasks runs highest. The top of that list:

• Data entry keyers. Most of the role is now in the model's range.
• Medical transcriptionists. Time-intensive transcription tasks have collapsed.
• Database architects. Substantial penetration on technical design tasks.
• Radiologists. AI hits high success rates on core knowledge work, though accountability still doesn't transfer.

Notice what that list isn't. It isn't truck drivers, retail workers, or service jobs. The most-exposed roles in 2026 are knowledge jobs with a structured input and a structured output. If your role looks like "documents go in, documents come out," AI has been built to do roughly that.

The sector-level data backs this up. Anthropic found that businesses in the information sector use AI at roughly ten times the rate of accommodation and food services (Anthropic, 2025). The exposure follows the adoption. The job categories where AI is being wired into workflows right now are the same categories where displacement is showing up first: software, marketing, administrative support, technical writing, financial analysis. The categories where adoption remains slow (healthcare delivery, skilled trades, in-person service work) are showing very little employment effect from AI so far.

Separate Stanford payroll research points the same direction. Early-career roles in software, marketing, and administrative support have contracted. Within firms, entry-level hiring in AI-exposed jobs declined about 13% relative to less-exposed jobs after the proliferation of LLMs (Stanford Digital Economy Lab, 2026). That number deserves a moment, because it's one of the clearest early labor-market signals we have.

Thirteen percent. Among young workers. In AI-exposed roles. Already.

The pattern is worth saying out loud. The roles getting hit first are the ones that are mostly the automatable tasks. A senior software engineer's job bundles automatable work (writing functions, fixing bugs) with non-automatable work (architecture decisions, mentoring, weighing tradeoffs with the business). A junior engineer's job is mostly the automatable half. So the senior keeps their role with new tools; the junior never gets hired in the first place.

The Deskilling Problem Nobody's Talking About

This is the finding I keep coming back to, because it inverts the optimistic story almost everyone tells about AI.

The reassuring narrative goes: AI handles the boring repetitive work, freeing you up for the interesting strategic work. The Anthropic data says the opposite is happening.

When Anthropic measured the average education level required for the tasks Claude is being used on, they found something striking. Claude tends to cover tasks requiring roughly one year more education than the economy-wide average: 14.4 years versus 13.2 (Anthropic, 2026). In other words, AI isn't picking off the low-skill tasks first. It's picking off the higher-skill ones.

Anthropic's own framing for this is "net deskilling." Their example: technical writers and travel agents are losing the higher-skill parts of their work while retaining the routine residue. The strategic thinking, the synthesis, the judgment call that justified your salary band: that's exactly the part the model is good at. The "make sure the printer has paper" half of the job stays.

I'll be honest about why this is so under-discussed. Every executive memo about AI talks about freeing humans for "higher-value work." It sounds great and almost nobody pushes back on it. But the data doesn't support it. The higher-value work is what's being absorbed first because the model trained on a lot of it.

If you want to feel the shape of this directly, picture a content marketing role from three years ago. The skill stack was strategy (positioning, narrative, voice), creative judgment (which angle, which hook), and execution (writing the post, formatting it, posting it). In 2026, AI can do reasonable execution unsupervised, decent creative variation with prompting, and increasingly competent strategic drafts. What's left for a human full-time hire? Mostly the parts the model wasn't trusted with: the brand-voice review, the final call, the meeting where someone has to defend the decision.

That's not "freed up for strategy." That's "kept around to approve things." It's still a job. It pays less.

Pick another role and the pattern repeats. A junior paralegal in 2022 spent most of the week on document review, discovery summaries, and case-law lookups. The reasoning-heavy parts of that work, the parts a paralegal would say justified the title, are exactly what AI tools have been built to do. The remainder is filing, scheduling, and chasing down signatures. A junior financial analyst in 2022 spent the week building models, writing memos, and synthesizing reports. The synthesis is now drafted by a model. The remainder is updating spreadsheets and following up on data requests. The optimistic story would call this "freed up for higher-order work." The data calls it net deskilling.

The professional consequence over a five-year horizon is the one that should worry people most. If the higher-skill tasks have been delegated to AI for years, the human stops developing those skills. You can't promote someone into a senior role they've never practiced. That's how a profession hollows out from the inside without a single layoff announcement.

What People Are Actually Using AI for at Work

Step back from the displacement question for a second and look at what AI is actually being used for. The picture clarifies a lot of the panic.

OpenAI Global Affairs' summary of OpenAI Economic Research found that 3.5% of all sampled ChatGPT conversations were about finding work or moving through a job application process, and that around 40 million US adults have used ChatGPT for some job-search task (OpenAI Global Affairs, 2025). Among those conversations, the breakdown is:

• Resume and cover letter preparation: ~33%
• Career exploration and research: ~30%
• Recruiter communication: ~20%

The remainder splits across interview prep, salary research, and decision support. Read that list once more. Almost nobody is using ChatGPT to automate someone else out of a job. They're using it on themselves, to get through the job market they're inside of.

The age split is interesting. OpenAI found that older workers tend to use AI to build on existing career capital: rewriting resumes, refining cover letters, drafting outreach. Younger workers use it to build new career capital: interview prep, position research, career strategy. If you're senior and not using AI to compound what you already know, you're conceding ground to people younger than you who are using it harder.

Anthropic's data on workplace adoption is the other half of the story. 40% of US employees report using AI at work, up from 20% in 2023 (Anthropic, 2025). On the company side, 9.7% of US firms had adopted AI by August 2025. Information-sector businesses use AI at roughly ten times the rate of accommodation and food services, which is why "is my job AI-safe" feels like a different question depending on which industry you're sitting in.

The honest summary: the people using AI most are not the people automating jobs. They're knowledge workers using it to stay competitive in roles that are slowly being redefined around them.

The Reliability Ceiling Most Coverage Skips

Here's a quiet thing that happened in early 2026 that almost nobody covered. Anthropic walked back their own productivity estimate.

An earlier version of their analysis suggested AI could add 1.8 percentage points to annual US labor productivity growth. The January 2026 report revised that down to 1.0 to 1.2 percentage points once they factored in how often the model actually succeeds at the tasks it's asked to do (Anthropic, 2026). That's a one-third haircut on the projected impact, from the company building the model.

The reason is in the success rates. Anthropic measured how often Claude completed tasks correctly by the human-education level the task typically requires:

• Tasks at a high-school level (12 years of education): ~70% success
• Tasks at a college level (16 years): ~66% success
• Tasks requiring five or more human hours of effort: ~50% success

And the collaboration mode matters as much as the task. Multi-turn, human-in-the-loop work on Claude.ai hits 67% success overall. Single-turn fully-automated API calls hit 49%. Translation: the closer AI gets to working alone, the less reliable it gets.

That ceiling is the most important factor in whether your specific job is AI-safe, and it's barely discussed. The jobs hardest to fully automate aren't the ones with the hardest tasks. They're the ones where errors are most expensive or hardest to detect. A radiologist's core image-reading task scores high on AI's capability list. Radiologist accountability still doesn't transfer to a model. The job persists because of who gets sued when it goes wrong.

The same logic protects roles you might not expect. Legal review. Compliance. Anything involving a customer relationship where trust is the asset. Anything where the error mode is "we lose the account." AI can do the underlying task; it can't carry the consequence.

If you're trying to decide whether your role is durable, ask a sharper question than "can AI do this." Ask: "if the AI did this and got it wrong, who's on the hook?" If the answer is "the AI vendor" or "nobody really," your role is exposed. If the answer is "a specific human with a name and a license and a relationship with the customer," your role is more durable than the headline writers think.

The Bottom Line

So is your job AI-safe? The honest answer is that nothing is fully AI-safe and nothing is fully gone. What you actually want to know is where you sit on a few specific axes the data is finally clear about. Here's how to read your own situation against the numbers.

If you're entry-level in a knowledge role. The roughly 13% within-firm decline Stanford documented for entry-level hiring in AI-exposed jobs is the trend you're inside of. Don't wait it out. Get visible on the parts of work AI can't take credit for: relationships, ownership of outcomes, customer-facing wins, anything where a specific person's judgment is what shipped the result. The work that builds a track record is the work that survives.

If you're early-to-mid-career in a role that gets touched by AI daily. Watch for the deskilling pattern. If the strategic and judgment parts of your job are getting absorbed by AI tools your team has rolled out, and your day-to-day is drifting toward routine review work, that's the signal to push back. Volunteer for the work that doesn't have a tool yet. Take on cross-functional projects. Move toward where the consequence lives.

If you manage a team. The deskilling risk is not abstract. If your direct reports are using AI for their highest-impact work and doing routine work themselves, you'll wake up in 18 months with a team that's lost the skill base it was supposed to be building. Be intentional about which work AI does and which work humans still need to learn.

If you're job searching right now. You're searching in a market where companies have AI screening on one side and 40 million people using AI on the other. Generic applications get crushed in the middle. The differentiator is no longer polish, which everyone has. It's specificity, evidence, and fit. See our breakdown of job searching in the age of AI and the AI-proofing playbook for the specific moves that still work.

The 2026 reality is not the 2023 apocalypse. It's quieter and more uneven, with sharp losses concentrated in early-career and routine knowledge work, and a deskilling effect that the optimistic narrative is still pretending not to see. Knowing where you sit in that pattern is the actual answer to whether your job is AI-safe.

Pick one task in your week that you'd lose to AI if you had to. Replace it with one you wouldn't.