AI narrows the performance gap between education groups
A sizable education gap remained in the subsequent test without AI. Lower-education participants did retain part of their gain. Durable learning was not tested.
AI RADAR · CHECKED DAILY
What new studies say about AI.
Every day our own AI system reviews new studies, puts them in context and works out what they mean for companies. With the study of the day, topic radar and study archive.
Study of the day
Mixed learning groups broaden teachers’ AI use
Standardized post-course difference relative to same-specialty groups. Both groups had the same curriculum, instructor and AI access. A higher value means more different reported applications.
Random group assignment; self-report. No measure of teaching quality, student outcomes or productivity.
Goulas, Megalokonomou & Sotirakopoulos · SSRN
The additional AI use was driven mainly by grading rubrics, which had been uncommon before training. The study measured self-reports immediately after the course, not teaching quality or productivity.
Our practical takeaway
Mix specialties in AI training. Then assess specific new uses and their quality; the study measures reported use, not teaching quality or productivity.
Published: 13 Oct 2026
- Region
- Greece
- Setting
- Nationwide AI training for practising teachers
- Sample
- 750 teachers · groups of five
- Method
- Preregistered randomized experiment · post-course self-report
AI Topic Radar
We compare new studies with earlier findings. See which patterns are supported and which questions remain open.
Topics at a glance
4 areas · 9 themesData and technologylast added 19 Oct 202617sources+0 / 14 days
People and organisation+1 added today70sources+13 / 14 days
Workflows and controllast added 01 Nov 202660sources+10 / 14 days
Value and scalinglast added 01 Nov 202639sources+5 / 14 days
Tap for context and evidence · A source may belong to several topics
How the topics connect
Explore connections between studies, themes and the evidence collected so far.
AI topics connect
No matching content found
Drag or select a point
Search & filter
100sources in total9themes14trends
Study archive
AI improves performance on more demanding IT tasks
The experiment shows better solutions on tasks beyond the curriculum. Immediate comprehension did not decline. Whether the capability lasts was not tested.
More speed is not deep transformation
The survey distinguishes faster work from deeply changing the business. It does not measure process times or productivity.
AI-supported work needs checking before handoff
The study captures expectations of colleagues. It does not show how often AI results are actually checked.
More planning improves AI writing but takes time
Goal-setting and task-breakdown support improved writing quality. Participants nevertheless rated the AI as less helpful.
AI creates new tasks for directing agents
The survey describes new activities, not a measured productivity gain.
AI agents need instructions and checking
Researchers observed work with coding agents. This time share describes prompting and checking, not wasted time or a measured productivity gain.
AI incidents reach the top of the organisation
A reported incident may involve a wrong AI decision or exposed data. The survey captures escalations, not control effectiveness.
More AI contributions need enough review time
The survey captures experiences in open-source projects. More contributions demonstrate neither higher quality nor measurable business value.
83% report no productivity impact from AI yet
No reported impact does not mean AI is ineffective. The survey captures perceptions, not independently measured changes.
Using AI does not mean embedding it in core operations
The study reports a selected US entrepreneur group’s own assessment. It does not measure business impact.
More speed with AI does not remove every bottleneck
AI-related effort also includes reviewing results and correcting prompts.
Using AI does not yet mean embedding it in daily work
The Scottish survey distinguishes individual AI applications from embedding AI across most of an organization. It measures reported use, not productivity or business performance.
At Anthropic, more Claude sessions focus on instructions and oversight
This includes plans for AI agents, reviews of their results and monitoring dashboards. Anthropic examines what employees use Claude for.
Few respondents report rules for stopping AI initiatives
The survey distinguishes rules before production from rules for AI already in use. It measures reported practices, not proven business success.
Organised team chats: more correct AI answers in the test
The study tested how an AI uses information from internal team chats. It found better answers, but does not establish time savings in everyday work.
AI can fail again on the next attempt
The test shows that one successful attempt is not enough to assess an AI system’s reliability.
Investigating an AI incident often takes hours or days
74% say it takes hours or days to retrace AI actions. A further 9% say longer or are unsure.
AI rollout speed and changed work are different things
64% of surveyed US CEOs see their AI pace as on track or ahead. Only 9% say agents are already fundamentally changing their business.
Financial returns are one part of AI success
80% report positive AI returns. Only 7% meet the combined criteria of demonstrated value, changed workflows and alignment between leaders and employees.
Many test AI agents; few report mature controls
68% of surveyed automotive companies use or test AI agents in product development. Only 5% describe their rules and controls as fully mature.
Employees less readily allow expertise to be shared with AI
In the experiment, about 77% allowed expertise to be shared with AI, compared with about 97% for sharing with people.
People need to detect errors without AI too
A theoretical model identifies a risk: when AI checks errors and people invest less in checking skills, they detect fewer errors during an AI outage.
An AI test does not say how long AI can work alone
The metric reports human completion time for software tasks AI solves in half its attempts. It does not measure reliable runtime without supervision.
Reported AI use can be exaggerated
48% of 1,000 surveyed US employees said they had at least occasionally exaggerated their AI use or knowledge.
Demand is growing for people who can build AI solutions
Demand for building AI tools, automating workflows and AI engineering rose 51% in Workday customer job postings from September 2025 to July 2026.
Longer AI texts need more checking
If an AI text contains more claims but the number checked stays fixed, the sample covers a smaller share. Errors can more easily remain undetected.
What would change without your AI?
42 of 91 respondents report AI in operation. Only 12 say removing it would require changes to costs or staffing.
AI project work needs a place in everyday work
Over 80% of the participating experts withdrew from AI innovation work at the law firm studied. Their extra work received little recognition in performance reviews or pay.
More AI training is only weakly linked to extra time saved
In a survey of 859 employees, AI training was associated with more frequent use. The association with reported time savings was weak.
A polished AI draft does not establish expert judgement
The analysis describes a problem: AI makes polished drafts easier to produce, making it harder to tell who understands the content and can assess errors.
A new AI model needs suitable workflows
Bain describes AI value as depending on data access, connected workflow steps and clear decision rights. A new model alone does not provide those conditions.
Similar AI accuracy, different amounts of human review
At similar AI accuracy, GPT-5.4 needs about 39% human review and Claude Sonnet 5 about 30% to meet the same reliability target under the same assumptions.
AI value needs to remain after review and rework
The conceptual paper proposes assessing the whole workflow: time saved alone is insufficient if review and rework consume the benefit.
AI also needs to find the right information
With preselected documents, AI reaches the required evidence in an estimated 90% of assessments, compared with about 75% using all accessible documents. The test shows how prepared information eases retrieval.
AI assessments need an expert cross-check
The methodological analysis shows that assessments of texts and documents can change with the model and instructions. Matching AI ratings can still share the same error.
Test the whole AI working system
The paper proposes testing the model, data access, retrieval, tools, safeguards and human oversight together. A bank pilot illustrates the approach in concrete workflows.
Better work with AI does not automatically mean learning without it
In a patent-lawyer field experiment, AI especially improved less experienced participants’ results. Lasting gains on later tasks without AI appeared mainly among experienced lawyers.
AI costs and value need to be visible together
Only 12% of 2,131 surveyed executives say they consistently link AI costs and value across their company.
Feeling more productive does not mean having less work
Heavy AI users in the employee survey report higher productivity, but no clear reduction in workload.
Supervising AI is additional work
Practitioner reports describe employees explaining tasks to AI agents, supplying information, checking results and correcting errors. This supervision requires time and skills.
Regular AI use is linked to learning and feedback
Learning, feedback and regular AI use occur together in a survey of 800 mostly medium-sized Indian companies. The survey does not establish which drives which.
A suitable task explains AI use only partly
Repeated US surveys find different AI use across similar work. Technical suitability alone does not fully explain the differences.
AI agents need controlled access and spending
Across 25 expert interviews, data access, costs, trust and company knowledge limit wider agent deployment. The report recommends controls on permissions, spending and actions.
AI value is also linked to knowledge of the business
The study links AI investment and productivity growth particularly in jobs combining AI skills with practical learning of company routines.
Explore further
Search studies, themes or practical questions.
Sources & context
The studies provide the findings. Our recommendations are interpretations of that evidence.
Diverse Peers and the Application of AI: Evidence from a Field Experiment
Goulas, Megalokonomou & Sotirakopoulos · SSRN · 13 Oct 2026
What is new?
New signal: Mixed specialties broaden reported AI applications after identical training
Open tension: More reported use does not establish better quality or productivity
Trend change: No established trend break
Last research check: 04 Nov 2026. Each source has its own publication date.
Explore the collection 100 sources · 14 trends
Knowledge map as links
45 earlier briefings. Lines show related topics, not cause and effect. Movement has no analytical meaning.
Earlier archive notes are labelled with their original language.
METHOD
How the AI Radar works
Selection
Every day our AI system looks for new, publicly available studies on AI in companies: research papers, field experiments and surveys. We include what is meaningful for companies.
Context
Each study is summarised and assigned to a topic. Every source shows publisher, date, method, sample and an assessment of the quality of evidence – including whether the work is peer-reviewed.
Implications
‘What we take from it’ is our assessment, kept separate from the study’s findings. We only use figures from the source and flag self-reported data. No vendor or product recommendations, no legal advice.
Sources
Every study links to the original. We do not reproduce charts or passages from the studies; we summarise in our own words.
WHAT DOES THIS MEAN FOR YOU?
Which of these findings affects your company?
We turn findings into concrete steps: which area first, which goal, what involvement from your team.
Book a conversation ↗