The Productivity Question Nobody Can Answer Yet
⚡ Quick Summary
- ✓ NBER Executive Survey: 89% of executives report no measurable productivity gains from 3 years of AI adoption.
- ✓ MIT Zero-Shot Disaster Modeling: Engineers create a method to simulate unprecedented catastrophe scenarios without historical training data.
- ✓ Invisible Email Hijacking: Forcepoint X-Labs demonstrates zero-size HTML font attacks manipulating AI email summarizers.
- ✓ Anthropic Unified Memory: Claude chat and Claude Cowork memory systems merged into a single cross-product context layer.
- ✓ The Real World Gap: Shipping capable frontier models remains far easier than extracting secure, measurable enterprise value.
After a month of record model releases and record enterprise spending, this week produced a much less comfortable headline: a major survey found almost no measurable productivity payoff from three years of AI adoption. Alongside it, a genuinely clever piece of disaster-modeling research, and a fresh warning about how easily AI summarizers can be hijacked.
The NBER survey nobody wanted
The National Bureau of Economic Research surveyed executives about three years of AI adoption at their firms, and the results were stark: more than 90% reported no effect on their own firm's employment, and 89% reported no effect on productivity.
Notably, job cuts attributed to AI haven't slowed even as the productivity gains that were supposed to justify them remain largely invisible in the data. It's a sharp counterpoint to the steady drumbeat of "AI is transforming the economy" headlines, and a reminder that enterprise adoption announcements and measurable output are two very different things.
MIT models disasters it's never seen
In a genuinely novel piece of research published in Nature Communications on August 20, MIT engineers described a method that generates worst-case disaster scenarios without training on any historical examples of the catastrophe being modeled.
That's notable because every existing risk-estimation tool has needed real historical disaster data to work from — this approach sidesteps that requirement entirely, which could matter a great deal for regions or scenario types where good historical data simply doesn't exist.
Invisible text, visible consequences
Security researchers at Forcepoint X-Labs demonstrated a sharp warning for anyone relying on AI email summarizers: they planted instructions inside an email using zero-size, white-on-white HTML text — invisible to a human reading it in Outlook, but fully intact in the content passed to the summarizing model.
The visible message ran 537 characters; the hidden payload brought the total passed to the model to 1,009 characters. Every one of ten test runs produced a manipulated summary — in one case moving an invoice deadline and deleting a name from the visible output. Forcepoint's recommendation is straightforward: summarizers should extract only user-visible content and treat email as fundamentally untrusted data, not as a trusted document to be faithfully condensed.
Anthropic merges memory across products
In a smaller but practically useful update, Anthropic merged the memory system behind Claude's chat interface and Claude Cowork, meaning context built up in one no longer stays siloed from the other — a small step, but one that matters for anyone using both products as part of the same workflow.
The bigger picture
Between the productivity survey and the email hijacking research, this week is a useful corrective to the relentless model-launch news cycle: shipping capable models is one problem. Getting reliable, secure, measurable value out of them in the real world remains a substantially harder one.
Written by Best AI Tool Editorial Team
We test, review, and curate the best AI tools, models, and industry updates for freelancers, developers, and creators.