AI Won't Replace You Overnight — But It's Getting Better Every Quarter

Sixty-five percent. That's how often AI can now complete a text-based work task that would take you three to four hours — up from about 50% just one year earlier. Fact — [MIT/arXiv, April 2026]

If that trajectory holds, most text-heavy jobs won't face a single dramatic "AI moment." They'll face a slow, relentless annual upgrade. And that distinction matters more than any headline about robots taking your job.

A new study from MIT — led by Neil Thompson and eight co-researchers — examined over 3,000 text-based tasks drawn from the U.S. Department of Labor's O*NET database and gathered 17,000+ evaluations from workers who actually perform those tasks. Fact — [arXiv:2604.01363] The paper's metaphor is vivid: are we watching "crashing waves" that suddenly wipe out certain jobs, or "rising tides" that gradually lift AI capability across nearly everything?

The answer, according to the data, is overwhelmingly the latter.

Rising Tides, Not Crashing Waves

The researchers found "little evidence of crashing waves" — narrow, sudden capability surges that eliminate specific tasks overnight. Instead, they documented "substantial evidence of rising tides": broad-based, continuous improvements across nearly all text-based work. Fact — [arXiv:2604.01363]

Here's what that looks like concretely. In the second quarter of 2024, large language models could handle roughly 50% of text-based tasks that would normally take a skilled human three to four hours. By the third quarter of 2025, that success rate had climbed to about 65%. Fact — [arXiv:2604.01363] That's a 15-percentage-point jump in just over a year — not for one narrow skill, but spread across thousands of different tasks.

The researchers project this trend forward: by 2029, AI could achieve 80% to 95% success rates on most text-related tasks at minimum quality thresholds. Estimate — [arXiv:2604.01363] Reaching near-perfect or human-superior quality would take several more years beyond that.

What This Means for Text-Heavy Occupations

If your job involves writing, analyzing, summarizing, translating, coding, or processing text-based information, this research directly concerns you. The rising tide doesn't discriminate much by occupation — it lifts broadly.

Consider data entry keyers. Much of their work involves structured text processing — exactly the kind of task where AI success rates are climbing fastest. Our data shows their AI exposure is already among the highest across all occupations.

Customer service representatives face a similar trajectory. Handling inquiries, drafting responses, routing issues — these are text-based tasks where LLMs have made consistent gains quarter over quarter.

For editors and technical writers, the picture is nuanced. AI can now draft and revise text competently, but the quality bar for professional editing and technical documentation remains high. The MIT study notes that near-perfect quality — the standard these roles demand — will take "several additional years" beyond the 2029 projections. Claim — [arXiv:2604.01363]

Translators sit at an interesting intersection. Machine translation has improved dramatically, but the study's focus on "minimum quality standards" versus professional-grade output matters enormously here. A 90% success rate at minimum quality still leaves a significant gap for nuanced, publication-ready translation.

Software developers, paralegals, accountants, and market research analysts all work in text-rich environments where AI's rising tide is unmistakable. But "being able to do the task" and "replacing the person who does the task" are very different things — a point the researchers emphasize.

The Adoption Gap: Capability vs. Reality

Perhaps the most important finding for anyone worried about their job: the researchers stress that adoption timelines may substantially exceed capability development timelines. Claim — [arXiv:2604.01363] Just because AI can do a task doesn't mean organizations will implement it.

Think about it this way. Spreadsheet software could automate many accounting tasks decades ago, but accountants didn't vanish — their work transformed. The same pattern is likely here. The MIT team explicitly notes that organizational implementation requires far more than technical capability: it needs process redesign, trust-building, compliance frameworks, and workforce adaptation.

This adoption gap is why the "rising tides" metaphor matters so much. A slow, steady rise gives workers, companies, and policymakers time to adapt. A crashing wave wouldn't.

What You Should Do With This Information

If you work in a text-heavy role, the data suggests three practical takeaways.

First, expect augmentation before replacement. The 65% success rate means AI is already a useful collaborator on many tasks, but it fails often enough that human oversight remains essential. Learn to work with AI tools now — that's where the immediate career advantage lives.

Second, invest in judgment and quality. The gap between "minimum quality" success (where AI is heading toward 80-95% by 2029) and "superior quality" performance (which remains years further out) is exactly where human expertise retains its premium. The tasks that require deep domain knowledge, ethical judgment, and stakeholder relationships are the last to be automated.

Third, watch the trend, not the snapshot. A 15-percentage-point annual improvement is significant. What AI couldn't do reliably last year, it may handle competently this year. Build a habit of reassessing which parts of your workflow could benefit from AI assistance every six months.

The MIT researchers have given us something valuable: a data-driven framework that replaces fear-mongering with measurement. The tides are rising, but they're not crashing waves. You have time — use it wisely.

Sources

• Mertens, M., Kuzee, A., Harris, B.S., Lyu, H., Li, W., Rosenfeld, J., Anto, M., Fleming, M., & Thompson, N. (2026). "Crashing Waves vs. Rising Tides: Preliminary Findings on AI Automation from Thousands of Worker Evaluations of Labor Market Tasks." arXiv:2604.01363. https://arxiv.org/abs/2604.01363

Update History

• 2026-04-04: Initial publication based on arXiv:2604.01363 (April 2026).

---

This analysis was produced with AI assistance (Claude claude-opus-4-6). All claims are tagged with evidence strength indicators and linked to source material. For detailed automation data on specific occupations mentioned, visit the linked occupation pages.