Wins and losses
A mid-month experiment
Just for fun I asked Claude Pro (on Sonnet 4.6) to round-up this month’s (so far) conversation from the AC Artificial Intelligence WhatsApp group, where comms and PR professionals trade research, news and arguments about AI’s effect on the profession and beyond.
Andrew Bruce Smith submits a lot of the content, supported by many others, for what I think has become a great source for news and information about AI in the comms world.
Here’s the playback from Claude. What do we think? Anything lost? Anything gained?
The efficiency-gain illusion
The month opened with a preprint that became one of its most-discussed pieces. Across three studies of nearly 2,700 people, researchers found that AI rarely saves time on small everyday tasks. People predicted they would use AI on a third of easy tasks but actually reached for it on closer to half, expected an average saving of 56 seconds but got only 7.5, and on the simplest jobs were measurably slower with AI than without it. Worse, using AI bred more confidence in its benefit rather than less, a feedback loop that entrenches the illusion.
For a comms audience, the group’s read was sharp: most of the daily texture of the job (rewriting a line, fixing spelling, summarising a short note) sits in exactly the zone where AI doesn’t help, while the genuine wins show up on heavier lifting. Sonya flagged the deeper worry, that this also threatens the small judgement calls that keep a practitioner’s hand in, and Andrew Bruce Smith connected it to the wider habit of treating AI as a single, undifferentiated tool, with most users still unclear on the difference between a chat assistant and a reasoning model.
AI search, SEO and the fight over attribution
A long-running theme was the reshaping of search by AI. One widely shared piece mapped classic SEO spam tactics onto their AI-era equivalents, from prompt injection standing in for cloaking to “vector neighbour poisoning” via expired domains. Its closing line stuck with the group: traditional search indexed the web, but AI search believes it, so whoever controls or poisons the underlying data pools gains outsized influence over what these systems confidently tell users.
Google’s Sundar Pichai said he is relaxed about AI Mode gradually displacing classic search, betting that subscriptions and advertising will sustain the business even as publishers face a “traffic cliff” of citations without clicks. In a partial counter, the UK’s Competition and Markets Authority confirmed that publishers can now opt out of Google’s AI Overviews, with a nine-month implementation window, while Google’s Search Console began surfacing AI citation data (though not referral traffic). DuckDuckGo, meanwhile, reported a real surge in installs from users actively rejecting AI-heavy search.
The professional body AMEC weighed in too, publishing GEO principles that pushed back on citation share becoming “the new vanity metric,” arguing the real question is whether a brand is discoverable, accurately represented and linked to outcomes.
Detection, disclosure and trust
A recurring strand concerned how badly AI detection tools work and what that does to trust. A Sydney academic’s Copilot-assisted opinion piece, written to warn students against using AI, was itself flagged by detection software and pulled by the Herald for non-disclosure, an episode the group found almost too on-the-nose. DocTim shared an anecdote from his team: a publication rejected a wholly human-written op-ed as AI-generated, and when the detection tool was run on South Africa’s 1955 Freedom Charter, it flagged that too as 60% AI-written.
Ethan Mollick’s reflections on writing his new book, Co-Existence, added a practitioner’s account of these tensions: heavy AI use for research, fact-checking and unblocking, but the prose itself written by hand because readers expect his voice and because, in his view, AI is still poor at sustained long-form narrative. He’s now also designing content explicitly for AI agents to read and recommend, what he calls “AIO” rather than SEO.
Compute, cost and the economics of scale
Several threads tracked the sheer scale of AI spending. Sam Altman noted that a token usage level once considered a global outlier six and a half years ago is now roughly the average user, while OpenAI’s heaviest internal user now consumes about 100 billion tokens a month. A Reddit account of Claude Code spawning 339 sub-agents in ten minutes, burning a Max plan’s quota in one task, did the rounds as a cautionary tale about parallel agent use. Andrew Bruce Smith’s own experiments with 20 parallel analyst agents drew a similar lesson: don’t default to multi-agent set-ups, and route cheaper models to bulk work while reserving frontier models for synthesis.
A Guardian piece pulled together the macro picture: AI-linked stocks now account for nearly half the S&P 500’s value, capital expenditure on AI infrastructure is forecast to more than double by 2031, and US GDP growth has leaned heavily on AI-related investment, all while only 7% of organisations using AI describe it as “fully deployed.” Teradata’s decision to skip 2026 pay rises to fund AI investment, alongside similar moves at TTEC, prompted debate about whether cutting compensation to fund AI is a real necessity or simply the path of least resistance.
Anthropic in the spotlight
Anthropic featured heavily this month: its own essay on recursive self-improvement (with internal claims that over 80% of code merged to its production codebase is now Claude-authored), a Financial Times profile examining the tension between its safety mission and its commercial trajectory toward a reported $1tn IPO, and its compute deal with xAI/SpaceX, which the group discussed as a striking role reversal of a one-time rival becoming an infrastructure supplier. A Claude-generated analysis of Anthropic’s own “AI-native startup” playbook document drew approval from Andrew for its self-aware disclosure of potential bias, and for treating the document as a specimen of content marketing as much as practical advice.
Jobs, juniors and the broken ladder
The health of the early-career pipeline was a constant undertone. One paper, shared by Elif, complicated the popular narrative that GenAI is driving down junior hiring, arguing that working-from-home is the stronger confound once both effects are modelled jointly. A separate employer survey found that nearly three times as many talent leaders expect AI to increase entry-level hiring as expect it to cut it, though AI literacy ranked dead last among the skills employers actually value. Against that, Bain’s “Next Monday” research argued teams are shrinking toward small, senior, AI-native pods, with apprenticeship inverting so juniors learn by verifying agent output rather than producing it themselves. James Crawford and Kerry Knight both raised the structural worry directly: what happens in three to five years if the junior pipeline isn’t maintained now.
The FT’s piece on AI disrupting the Big Four and strategy consultancies drew an unusually rich 226-comment thread, which the group’s own AI-assisted analysis distilled into several camps: clients increasingly doing the work themselves, scepticism that AI-native boutiques are any more durable than incumbents, and a recurring fear that the foundation labs (Anthropic, OpenAI) could ultimately disintermediate the consultants altogether.
Risk, extremism and harder edges
A heavier thread tracked the darker fallout from rapid AI deployment. A Guardian feature documented a rise in anti-AI political violence, from an attempted arson at OpenAI’s headquarters to a fatal attack in San Diego citing “AI slop,” and argued AI has become a unifying fixation across otherwise disparate extremist groups. Separately, a report on a China-based cybercrime network found it had used Gemini to generate convincing phishing sites at scale, stealing an estimated 3.87 million credit card numbers since 2023. Locally, the group discussed the difficulty of correcting AI-generated misinformation about named individuals, prompted by a case of a former police officer wrongly linked online to a high-profile arrest.
Culture and practice
Lighter threads ran throughout: AI-generated event posters for village fairs versus the encroachment of “amateur land” briefs into commercial design work; Katie Parrott’s system of persona-driven AI editorial skills (Sorkin for momentum, Mom for accessibility, Asshole for stress-testing arguments); and Antony Mayfield’s framing of “agents as loops,” with Estonia’s schools curriculum reform held up as a rare example of an institution making a clear, deliberate choice about which cognitive work to hand to machines and which to protect.
As ever, the group’s own habit of running AI analysis on AI-related news, sometimes multiple models in parallel for a “council” verdict, continued to be both a working method and a running joke about the layers of self-reference involved in covering this beat.
For Further Reading
AI search spam and the new SEO: https://www.iloveseo.net/spam-in-the-age-of-ai-search/
Pichai comfortable with AI Mode replacing classic search: https://www.searchenginejournal.com/googles-ceo-is-comfortable-with-ai-mode-replacing-classic-search/577923/
CMA lets UK publishers opt out of Google AI Overviews: https://www.bbc.co.uk/news/articles/c775pp26yz5o
DuckDuckGo installs up as users reject AI-heavy search: https://techcrunch.com/2026/05/26/duckduckgo-installs-are-up-30-as-users-reject-being-force-fed-googles-ai-search/
AMEC GEO Principles: https://amecorg.com/amec-geo-principles/
Sydney academic’s AI-assisted op-ed against AI use, flagged by detection software: https://www.theguardian.com/australia-news/2026/jun/03/sydney-academic-used-ai-opinion-piece-urging-students-to-avoid-using-it-ntwnfb
Ethan Mollick on writing Co-Existence with and without AI:
Sam Altman on OpenAI’s top token spenders: https://www.businessinsider.com/sam-altman-openai-top-token-spender-ai-costs-issue-2026-6
Claude Code spawning 339 sub-agents (Reddit): https://www.reddit.com/r/ClaudeAI/comments/1u5d02w/til_claude_code_spawned_339_subagents_from_a/
The AI boom explained in six charts (Guardian): https://www.theguardian.com/technology/2026/jun/07/billions-spent-hypothetical-returns-the-ai-boom-explained-with-six-charts
Teradata skips 2026 pay rises to fund AI investment: https://www.businessinsider.com/teradata-pauses-raises-employee-compensation-ai-budget-2026-6
Anthropic, “When AI builds itself” (recursive self-improvement): https://www.anthropic.com/institute/recursive-self-improvement
FT profile of Anthropic’s safety-versus-commerce tension: https://www.ft.com/content/e17665ea-c5ca-428a-839c-be5c1eacc35c
xAI’s pivot to becoming Anthropic’s compute supplier: https://www.theinformation.com/articles/xai-went-chasing-anthropic-powering
Is GenAI replacing junior workers, or is it WFH? (SSRN paper): https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6787638
Entry-level hiring in the AI era, employer survey: https://www.strada.org/news-insights/entry-level-hiring-in-the-ai-era-what-employers-are-thinking-and-doing
Bain, “Next Monday” moves for leaders on AI: https://www.bain.com/insights/the-2026-ceo-agenda-where-ambition-outpaces-execution/
FT on AI disrupting the Big Four and strategy consultancies: https://www.ft.com/content/d82d2a5c-74ab-4eb9-a658-fd5467e71670
Anti-AI tech extremism and political violence (Guardian): https://www.theguardian.com/technology/2026/jun/07/anti-ai-tech-extremism-violence
Chinese cybercrime network’s use of Gemini for phishing: https://www.helpnetsecurity.com/2026/06/12/google-china-based-cybercrime-network-lawsuit/
Former officer wrongly linked online to the Henry Nowak arrest: https://www.theguardian.com/uk-news/2026/jun/03/former-officer-hampshire-hiding-after-being-falsely-linked-henry-nowak-arrest
AI poster “slop” and local events flyers (Independent): https://www.independent.co.uk/life-style/ai-poster-slop-local-events-flyer-b2989792.html
Katie Parrott’s AI editorial skills and guardrails (Every): https://every.to/working-overtime/my-editor-caught-me-sounding-like-ai-now-ai-catches-me-first
Antony Mayfield, “Agents are loops”:
Antony Mayfield on Bloom’s taxonomy and Estonia’s AI Leap:
