We were breathless, bullish, and—let’s be honest—way too sure. Here are five 2020-era AI takes that didn’t survive contact with reality, checked against today’s facts for you!
Did you know these claims?
Back in 2020, we cheered for robotaxis, braced for mass job loss, and trusted flashy COVID forecasts. Five years on, the data tells a very different story:
1) “Robotaxis are imminent—full autonomy is basically done.”
What we said in 2020
Elon Musk had already vowed “over a million robotaxis” by 2020 and said Tesla was “very close” to Level 5 autonomy.
What actually happened by 2025
Tesla’s “Full Self-Driving” remains a supervised driver-assistance system (effectively Level 2), while genuine driverless services are limited to specific zones and pilots—led mainly by Waymo in a handful of U.S. cities, with international plans still staged and regulated. In short: impressive progress, not a solved problem.
A bold promise met a much messier road.
2) “AI will replace most jobs by 2025.”
What we said in 2020
Doomy forecasts went viral: mass unemployment was “inevitable.”
What actually happened by 2025
Evidence shows high exposure but mixed outcomes: the IMF estimates ~40% of global employment is exposed to AI, with advanced economies more affected—yet much of that exposure is complementary, not replacement.
Employer surveys (WEF) also projected both creation and loss, with a net −14M jobs by 2027 (69M created, 83M eliminated), i.e., churn and reskilling—not a clean sweep.
The jobs story turned out to be transformation, but not total annihilation.
3) “COVID proved AI can forecast complex crises at scale.”
What we said in 2020
Hype suggested ML would predict, diagnose, and triage COVID-19 with uncanny accuracy.
What actually happened by 2025
A living BMJ review found most early COVID prediction models were at high risk of bias and not fit for clinical use; generalization across hospitals and populations was a persistent Achilles’ heel. Post-mortems in top journals echoed the same concerns about data quality, bias, and reproducibility.
So it didn’t work—at least not reliably back then.
4) “Bias? Fixed. Facial recognition and ‘emotion AI’ are ready for prime time.”
What we said in 2020
Vendors touted major accuracy gains and even claimed systems could read emotions from faces.
What actually happened by 2025
NIST’s demographic evaluations and a string of wrongful arrests showed real-world harms and uneven performance. Meanwhile, a major scientific consensus cautioned that you cannot reliably infer internal emotional states from facial movements alone—undercutting the premise of “emotion AI.” Europe went further: an EU AI Act now prohibits emotion recognition in workplaces and schools and tightly curbs biometric uses.
In the end: Not even close—and regulators noticed.
5) “Regulation will lag for a decade—nothing big is coming soon.”
What we said in 2020
The common refrain: by the time lawmakers act, the AI train will be long gone.
What actually happened by 2025
In China arrived a binding, nationwide regime for public-facing GenAI via the Interim Measures for Generative AI Services (effective 15 Aug 2023), laying out provider duties on data, content safety, and transparency—part of a broader stack regulating recommender algorithms and deep synthesis.
Abu Dhabi set up the Artificial Intelligence & Advanced Technology Council (AIATC) in Jan 2024 and reconstituted it on Feb 12, 2025 to steer AI policy and oversight. It free zones moved with guidance and licensing (e.g., DIFC practical note on using LLMs in proceedings.)
An EU AI Act was published in the Official Journal on 12 July 2024 and entered into force weeks later, with staggered obligations rolling out through 2025-2027—including outright bans on certain practices (like emotion recognition in specific contexts). Whatever you think of it, this is not “no regulation.”
Regulation arrived faster and tougher than many expected.
The Pivot: Why AI Agents Are The Future
After five years of hype whiplash, the winners aren’t vague promises—they’re AI agents that do work, end-to-end, inside clear guardrails. Think task-grounded, tool-using systems that plan, execute, and verify against business-grade metrics.
Agents don’t need sci-fi autonomy to deliver value; they need real workflows, observability, and a human in the loop. That’s where 2026 is heading: from predictions that age badly to agentic execution that ships results. Especially for businesses, no matter which scale.
Got interested?
