AI-Moderated Discussions Create 'Illusion of Inclusion,' Study Finds
May 17, 2026
D.A.D. today covers 14 stories from 4 sources. What's New, What's Innovative, What's Controversial, What's in the Lab, and What's in Academe.
D.A.D. Joke of the Day: My AI passed the bar exam, medical boards, and the CPA test. It also told me Abraham Lincoln invented WiFi. So I'd say it's ready for middle management.
What's New
AI developments from the last 24 hours
Malta Will Offer Free ChatGPT Plus to All Citizens Who Complete AI Course
OpenAI and Malta announced what they call a 'world's first' partnership to provide ChatGPT Plus access to all Maltese citizens. The program requires completing an AI literacy course developed by the University of Malta, after which citizens receive one year of free access. Malta's small population (roughly 500,000) makes this feasible as a pilot, but the structure—pairing access with education—offers a template other governments could adapt.
Why it matters: This signals governments are starting to treat AI access as public infrastructure rather than just a consumer product—a shift that could shape how AI tools get distributed and regulated at the national level.
Discuss on Hacker News · Source: openai.com
Rust-Based Coding Agent Claims Fraction of Competitors' Memory Use
Zerostack, a new coding agent written in Rust, claims a RAM footprint of roughly 8-12MB—a fraction of existing tools. Community reaction on Hacker News has been enthusiastic, with developers citing frustration over alternatives like Claude Code (multiple gigabytes) and opencode (which some report leaking up to 6GB of memory). One commenter reported reviewing the codebase and found nothing concerning. Others are requesting performance benchmarks comparing it to established agents.
Why it matters: For teams running AI coding assistants on developer machines or resource-constrained environments, memory efficiency could determine which tools are practical to keep running.
Discuss on Hacker News · Source: crates.io
Open-Source Video Generator Promises One-Minute 720p Output
SANA-WM, a new open-source AI model for generating 1-minute 720p video, has been announced at 2.6 billion parameters—compact enough to potentially run on consumer hardware. Early community reaction is skeptical: users report the download is currently unavailable, question whether it will actually fit on a 24GB GPU like the RTX 4090, and note the demo outputs have visual consistency issues and a video-game-like quality that suggests synthetic training data. Comparisons to closed-source competitors like Seedance and Kling have been unfavorable.
Why it matters: Open-source video generation remains far behind commercial offerings, and this release—while technically ambitious—illustrates the gap that limited training data creates.
Discuss on Hacker News · Source: nvlabs.github.io
Frontier Models Now Solve Most Hacking Challenges, Researcher Claims
A competitive security researcher argues that frontier AI models have effectively broken open Capture The Flag competitions—the hacking challenges used to train and evaluate cybersecurity talent. The claim: top models can now solve most challenges with minimal human input, turning what was a test of security skills into a measure of who can spend the most on AI tokens. The author says GPT-4 could already one-shot many medium-difficulty challenges, while newer models allegedly handle even 'insane difficulty' exploits that once required deep expertise.
Why it matters: If accurate, this signals both a credentialing crisis for cybersecurity hiring (CTF wins have long been résumé gold) and a preview of how AI may reshape other skill-based competitions and assessments.
Discuss on Hacker News · Source: kabir.au
Developer Reportedly Spent $1.3 Million on OpenAI Tokens in 30 Days
A single developer reportedly spent $1.3 million on OpenAI API tokens in 30 days—roughly $19,000 daily—to build OpenClaw, burning through an estimated 600 billion tokens. The developer was subsequently hired by OpenAI. Community reaction has been skeptical: one commenter questioned whether the spending made him five times more productive than a $200,000/year developer and called it environmentally harmful. Others noted the project still required significant hands-on management despite the massive spend. Some calculated that alternative approaches could have cut costs by 100x.
Why it matters: This is an extreme data point in the emerging question of AI tool economics—whether throwing compute at problems delivers proportional value, and what sustainable AI-assisted development actually looks like.
Discuss on Hacker News · Source: twitter.com
What's Controversial
Stories sparking genuine backlash, policy fights, or heated disagreement in the AI community
AI-Exposed Jobs Show First Signs of Decline, But Effect Remains Modest
Bureau of Labor Statistics data shows 18 occupations flagged as AI-exposed—roughly 10 million jobs—saw employment drop 0.2% between May 2024 and May 2025, while overall US employment grew 0.8%. The gap is modest but directionally notable. BLS projections still list software-related roles among the top 15 growth categories through 2033, suggesting any displacement effect remains limited. Community reaction has been skeptical, with some arguing the losses reflect broader economic pressures rather than AI specifically.
Why it matters: This is early, ambiguous data—but it's the kind of signal executives and policymakers will watch closely as they debate whether AI displacement is real or overstated.
Discuss on Hacker News · Source: bloomberg.com
What's in the Lab
New announcements from major AI labs
OpenAI Pitches Codex for Executive Briefs, Not Just Code
OpenAI published a guide positioning Codex as a tool for business operations teams—not just developers. The use case: pulling scattered information from project trackers, dashboards, meeting notes, and planning docs to generate first drafts of executive briefs, strategic updates, and leadership decision packets. OpenAI claims this speeds up the grunt work of synthesizing operational data while humans retain judgment on recommendations. No benchmarks or case studies accompanied the guide.
Why it matters: This signals OpenAI pushing Codex beyond its original coding-assistant identity toward general knowledge work—a direct play for the same enterprise productivity market Claude and Gemini are targeting.
Codex Guide Shows Data Teams a Draft-First Workflow for Reports
OpenAI published a guide for data science teams using Codex to convert raw inputs—questions, dashboards, data files—into polished analysis deliverables. The company says Codex can generate first drafts of root-cause briefs, impact readouts, and KPI memos complete with charts, caveats, and source links. Teams would then validate and refine the output. No performance data or case studies were provided—this is vendor guidance, not independent testing.
Why it matters: If the workflow holds up in practice, it could compress the time between 'here's a data question' and 'here's a shareable answer'—worth testing if your analysts spend significant time on recurring report formats.
Sales Teams Get OpenAI's Pitch: Codex as Research Assistant
OpenAI published a guide showing how sales teams can use Codex to aggregate data from CRMs, call notes, emails, and Slack to draft pipeline briefs, meeting prep packets, and account plans. The workflow positions Codex as a research-and-drafting layer—pulling scattered context into working documents while sellers retain strategic judgment. No performance data or customer results were provided.
Why it matters: This is OpenAI making a direct pitch for enterprise sales workflows, signaling the company sees administrative grunt work—not just coding—as Codex's growth market.
What's in Academe
New papers on AI and its effects from researchers
How Open-Weight Models Are Cutting the Cost of Long Documents
A technical analysis examines how recent open-weight models—from Google's Gemma 4 to DeepSeek V4—are implementing new architectural techniques to reduce the computational cost of processing long documents. The approaches include KV sharing (reusing cached data across attention layers), multi-head clustering, and compressed attention mechanisms. These are engineering optimizations that make long-context AI cheaper and faster to run. The article provides no benchmark data or performance comparisons.
Why it matters: This is infrastructure-level progress—if these techniques mature, expect longer context windows and lower API costs from providers building on open-weight models.
Source: magazine.sebastianraschka.com
Walking Coach Study Suggests Emotional AI Beats Pure Information
Researchers built SmartWalkCoach, a mobile AI system that acts as a walking companion using three lightweight agents: one for route planning, one for motivational prompts during walks, and one for post-walk reflection. In a small field study (12 participants), adding motivational dialogue to walking guidance significantly improved users' positive feelings and experience compared to information-only directions. The finding suggests that AI assistants benefit from emotional engagement, not just functional accuracy—a design principle that could extend beyond fitness apps.
Why it matters: For anyone building or evaluating AI assistants, this is early evidence that companion-style interaction outperforms purely informational delivery—relevant as voice assistants and workplace copilots compete on user experience, not just capability.
Grading AI Desktop Agents on a Curve, Not Pass-Fail
Researchers propose a new approach to evaluating AI agents that operate graphical interfaces—the kind that click buttons, fill forms, and navigate apps on your behalf. Current systems judge each action as simply right or wrong. The new method, BBCritic, instead measures how close an action is to correct on a continuous scale, similar to how a GPS shows you're 50 feet from your destination rather than just 'not there yet.' The smaller 3B-parameter model outperformed larger 7B competitors, suggesting the framing matters more than raw size.
Why it matters: As AI assistants that control your screen become more common, better evaluation methods could help developers catch near-misses before they become costly errors—clicking the wrong 'Submit' button is worse than hovering near the right one.
AI-Moderated Discussions Create 'Illusion of Inclusion,' Study Finds
A study of 879 participants making real charitable donations found that AI facilitation of group discussions didn't actually improve consensus—but did create measurable risks. When LLMs moderated small-group deliberations over $7,200 in charity allocations, they shifted some donation distributions by up to 5.5 percentage points. More troubling: participants reported feeling more included and trusted the process more in conditions where the AI exerted greater directional influence, even though actual participation equity didn't improve.
Why it matters: As organizations experiment with AI-facilitated meetings and deliberation tools, this suggests a specific governance concern: AI moderators may subtly steer outcomes while making participants feel better about the process than they should.
Citations Make AI Seem Trustworthy, but Users Rarely Check Them
A study of 20 users examining Microsoft Edge's Copilot found that adding citations made AI answers feel more trustworthy—but participants rarely bothered to check them. When they did fact-check, they often reached for the same sources the AI had already used, creating a circular validation loop. The research highlights how interface design choices shape trust in ways that may not correspond to actual accuracy.
Why it matters: As AI assistants become standard in browsers and productivity tools, understanding how design elements like citations influence user behavior—sometimes creating false confidence—becomes critical for both product teams and the professionals relying on these tools.
What's On The Pod
Some new podcast episodes
The Cognitive Revolution — Three Kinds of Software Survive: Tasklet's Andrew Lee on Competing to be a Horizontal Platform