Field note · AI Safety, Security & Adoption

The AI Reckoning.

Everyone moved fast. Now we're counting the cost. Safety culture collapsed at labs, companies replaced workers then quietly hired them back, and markets added trillions to companies building tools that researchers warn we don't yet know how to control. Here's what the data actually says.

0
Jobs explicitly attributed to AI in 2025 layoff announcements a 12× increase in two years
Challenger, Gray & Christmas · 2025
0
NVIDIA market cap as of May 2026 from $100B in 2019 to the world's most valuable company
TradingKey · May 2026
0
Of organisations now using AI regularly in at least one business function, up from 50% in 2022
Vention / McKinsey · 2025
0
Major AI companies with an adequate plan to control systems smarter than humans, per the Future of Life Institute
FLI AI Safety Index · Summer 2025
Safety culture and processes have taken a backseat to shiny products.
Jan Leike, former Head of Superalignment at OpenAI on resigning, May 2024
§ The adoption curve · With and without safety

What deployment actually looks like.

AI adoption accelerated faster than safety frameworks could keep pace. The gap between organisations deploying AI and those with formal AI risk governance is still wide.

Enterprise AI adoption

% of large organisations with active AI deployment · IBM / McKinsey / Vention
2020
40%
2021
44%
2022
50%
2023
55%
2024
80%
2025
88%

Formal AI risk governance in place

% of AI-deploying organisations with documented AI risk frameworks · McKinsey / MIT
2020
12%
2021
16%
2022
21%
2023
28%
2024
34%
2025
~38%

The gap between deployment and governance represents the structural risk the industry is sitting on. Only ~1 in 3 AI-deploying organisations has a documented framework for managing that risk. Sources: IBM Global AI Adoption Index 2023 · McKinsey State of AI 2024 · Vention AI Adoption Statistics Q1 2026 · MIT Sloan Management Review

§ Who's doing it right · The FLI Safety Index

Three approaches to safety. Three very different results.

The Future of Life Institute's 2025 AI Safety Index scored labs across transparency, preparedness, governance, and alignment research. The picture is not flattering but the gaps between labs are stark.

⬆ Ranked 1st
Anthropic
Public Benefit Corporation · Founded by ex-OpenAI safety researchers
TransparencyStrong
Alignment researchWorld-leading
Governance structureBest-in-class
External oversightGood
Founded 2021 by Dario Amodei and others who left OpenAI over safety concerns safety is the founding thesis, not an afterthought
Operates as a PBC with a Long-Term Benefit Trust; board is legally required to prioritise society over profit
Constitutional AI: safety rules are embedded in the model's training, not applied as post-hoc filters
Responsible Scaling Policy (RSP) defines enforceable thresholds (ASL-1 through ASL-4) that pause training if crossed
Does not train on user data; excelled in FLI bio-risk trials
→ Ranked 2nd · Declining
OpenAI
For-profit Corp (restructured 2025) · Backed by Microsoft
TransparencyModerate
Alignment researchDissolved 2024
Governance structureContested
External oversightLimited
Superalignment team formed July 2023, promised 20% of compute was dissolved May 2024 after both its co-leaders resigned over resource starvation
Jan Leike (Superalignment lead) publicly stated "safety culture has taken a backseat to shiny products" on exit
Published whistleblowing policy; only lab to do so; detailed pre-mitigation risk assessment
Converted from non-profit to public benefit corporation in 2025 amid significant board and governance drama
Did not publish a system card for GPT-4.1; released models before completing safety documentation
↓ Ranked Last · Critical
xAI (Grok)
For-profit · Founded by Elon Musk · Merged with X
TransparencyNone
Alignment researchMinimal
Safety governanceDismantled
External oversightAbsent
Released Grok 4 in July 2025 without any publicly disclosed safety report breaking commitments signed at the Seoul AI Safety Summit (May 2024)
Grok called itself "MechaHitler" in July 2025 after system prompt modification; went offline; relaunched days later with Grok 4
Generated sexualised images of children in January 2026 after safeguards were bypassed; small safety team had lost several members weeks prior
OpenAI and Anthropic researchers publicly called the approach "reckless" and "completely irresponsible" Harvard/OpenAI researcher Boaz Barak, July 2025
Grok 4.1 Fast rated the most dangerous model in a Stanford study on delusional spirals, reinforcing paranoid and grandiose beliefs in users

Source: Future of Life Institute, AI Safety Index Summer 2025 · Stanford Center for Research on Foundation Models · arXiv 2502.09288 · International AI Safety Report 2025 (Yoshua Bengio et al., arXiv:2510.13653)

§ The landscape · Who's building what

Every major player, mapped.

The AI industry is not one monolith it's a set of very different bets on where intelligence is going, who controls it, and what it's for. Here are the companies that define the field, with their safety posture alongside their scale.

Filter
Frontier models · US
OpenAI
$300B+
Valuation · 2025
Maker of ChatGPT and the GPT model family. First mover in consumer generative AI. Backed by Microsoft ($13B). ChatGPT reached 100 million users in 2 months the fastest consumer product adoption in history. Now a public benefit corporation after years of non-profit governance drama.
⚠ Safety drift
Frontier models · US
Anthropic
$61B
Valuation · 2025
Maker of Claude. Founded 2021 by ex-OpenAI researchers who left over safety concerns. Operates as a Public Benefit Corporation with legally binding safety obligations. Ranked #1 in the FLI AI Safety Index. Constitutional AI is its core architectural differentiator safety baked in, not bolted on.
✓ Safety leader
Frontier models · US
Google DeepMind
$4.2T
Alphabet market cap
The merger of Google Brain and DeepMind. Makers of Gemini. Integrating AI across Search, Workspace, and Cloud. Academic in approach, requiring safety researchers to flag capability-related risks early in model training. Ranked 3rd in FLI Safety Index. Gemini's 2024 image generation debacle showed even disciplined labs can fail at scale.
Structured approach
Frontier models · US
xAI (Grok)
$50B
Valuation · 2025
Elon Musk's AI lab, merged with X (Twitter) in 2024. Grok is embedded into one of the world's largest social platforms with 600M+ users. Released Grok 4 in 2025 without a safety report. Generated CSAM in January 2026. Ranked last in every FLI Safety Index category. Musk has internally pushed back against guardrails.
✗ Critical risk
Open source · China
DeepSeek
$10B
Valuation · 2026 (raising)
The industry's biggest shock of 2025. Released R1 in January 2025 under the MIT license matching OpenAI o1 performance at a claimed training cost of $6M vs GPT-4's $100M. Within a week it topped the US App Store. Wiped $600B from NVIDIA's market cap in a single day. Built using capped export chips under US sanctions. A genuine "Sputnik moment" for AI.
Open weights · MIT
Open source · US
Meta AI (Llama)
$1.7T
Meta market cap
Meta's Llama family is the most widely deployed open-weight frontier model series. Llama 3 and 4 are used in research, enterprise, and consumer products globally. Meta's open-source strategy is partly competitive keeping frontier AI accessible limits OpenAI's moat and powers Meta's own products at scale. Adopted mixture-of-experts patterns similar to DeepSeek after R1's release.
Open weights
Open source · France
Mistral AI
$13–14B
Valuation · Sep 2025
Europe's flagship AI lab. Paris-based, founded 2023, total funding ~€3B. Open-weight models including Mistral 7B, Mixtral (MoE), and the Le Chat assistant. Represents Europe's strategic bet on AI sovereignty building frontier models that don't rely on US labs. Partners with the EU on AI Act compliance and sovereign cloud deployments.
Open weights · EU-native
Voice AI · UK / Poland
ElevenLabs
$11B
Valuation · Feb 2026
The leading AI voice synthesis platform. Founded 2022. Revenue grew from $25M to $90M ARR in 2024 alone. Backed by NVIDIA, Sequoia, and a16z. Raised $500M in Feb 2026 eyeing an IPO. Offers ultra-low-latency speech (<75ms), 32 languages, and a voice marketplace where actors earn royalties. Selected for Disney Accelerator 2024. The voice layer of the generative AI stack.
Application · Audio AI
AI Search · US
Perplexity AI
$18–20B
Valuation · 2025
The "answer engine" challenging Google Search. Founded 2022. 780 million queries in May 2025 alone up 240% in 9 months. Backed by NVIDIA, Bezos, SoftBank, IVP. Nearly $100M ARR. Launched the Comet AI browser in July 2025 with autonomous browsing, task completion, and multi-model switching between Claude, GPT-4o and Mistral. 45M+ monthly active users as of early 2026.
Application · Search
Open-source hub · France / US
Hugging Face
$4.5B
Valuation · 2024
The GitHub of AI. Hosts 500,000+ models, 100,000+ datasets, and is the default open-source platform for the entire AI research community. Used by Google, Meta, Microsoft, and Amazon. Total funding $400M. DeepSeek R1 derivatives alone accumulated 10M+ downloads on Hugging Face. Backed by Google, NVIDIA, AWS, IBM, Salesforce, and others.
Open ecosystem
Image generation · US
Midjourney
$50M+
ARR · Bootstrapped
The dominant AI image generation platform and one of the few profitable AI consumer companies. Entirely bootstrapped, no venture capital. Started on Discord; expanded to a web app in 2024. Part of the leading group of 18 AI-native firms with $50M+ ARR. Faces ongoing legal challenges over training data from artists and copyright holders a defining tension for generative AI image models.
Copyright disputes ongoing
Enterprise AI · Canada
Cohere
$5B+
Valuation · 2025
Founded by Google Brain alumni, Cohere targets enterprise and sovereign AI not consumer chatbots. Hit $100M ARR in 2025. Backed by NVIDIA, Salesforce, and Oracle. Used by McKinsey, Salesforce, and LivePerson. Specialises in retrieval-augmented generation (RAG) and on-premise deployments for regulated industries. Canada's AI champion and a quiet giant in enterprise LLMs.
Enterprise focus · Compliance-friendly
AI Infrastructure · US
Scale AI
$13.8B
Valuation · 2024
The data labelling and AI evaluation company behind most major model training pipelines. OpenAI, Anthropic, Google, and the US Department of Defense are all customers. CEO Alexandr Wang the world's youngest self-made billionaire at the time of Scale's first valuation. The unglamorous infrastructure layer that all frontier AI depends on: human-labelled data at scale.
Infrastructure · Data
Image / open source · UK
Stability AI
Restructured
Post-2024 rebrand
Maker of Stable Diffusion the open-source image model that democratised AI art. A cautionary tale on governance: founder Emad Mostaque resigned in March 2024 amid reports of financial mismanagement, unpaid bills, and investor disputes. The company restructured under new leadership. Demonstrates that "open source" doesn't automatically mean sustainable or well-governed.
Governance failure · Restructured
AI platforms · US (Defence)
Palantir
$290B+
Market cap · 2025
The controversial data analytics and AI platform used by governments and militaries. AIP (AI Platform) brought LLMs into defence, intelligence, and enterprise operations. Stock surged over 300% in 2024 on AI momentum. Raises persistent ethical questions about AI in warfare and surveillance the most acute real-world manifestation of AI safety applied to high-stakes, life-and-death decisions.
AI in defence · Contested
AI chips · US
Cerebras Systems
$8.1B
Valuation · Sep 2025
The NVIDIA challenger building wafer-scale AI chips the CS-3 chip is the size of a dinner plate and contains 4 trillion transistors. Claims 20× faster inference than GPU clusters for certain workloads. Backed by $2.8B in funding. The bet that AI chips don't have to look like NVIDIA GPUs and that inference speed, not training, will be the next hardware arms race.
Infrastructure · Chips

Valuations: CNBC (ElevenLabs Feb 2026) · Wikipedia (DeepSeek) · Vestbee (Mistral Sep 2025) · Winbuzzer (Perplexity Jul 2025) · TradingKey (Alphabet/NVIDIA May 2026) · Wellows AI Startups 2026 · FLI AI Safety Index (safety ratings)

§ The record · 2020–2026

When it happened, what it meant.

The AI era in sequence from the foundational research, through the product arms race, to the safety collapses and regulatory responses. Filter by category.

Filter
Sep 2021
Safety win
Anthropic founded by Dario Amodei, Daniela Amodei & seven other OpenAI researchers
The founding thesis: AI safety is not a feature to add later. The group left explicitly over concerns that OpenAI was prioritising speed over alignment research.
Nov 2022
Market
ChatGPT launches. 1 million users in 5 days. 100 million in 2 months.
The fastest-adopted consumer product in history triggered the generative AI gold rush. NVIDIA, Microsoft, and every enterprise software company reset their roadmaps overnight.
Jan 2023
Market
Microsoft invests $10 billion in OpenAI. Integrates GPT across the entire product suite.
Microsoft's P/E ratio jumped sharply. The bet catalysed a wave of AI-first positioning across enterprise software real or otherwise.
Feb 2023
Failure
IBM freezes hiring for 7,800 roles it expects AI to replace within 5 years
CEO Arvind Krishna cited back-office and administrative functions as primary targets. The announcement made headlines as the first major corporate signal of AI-driven downsizing. Many roles were later quietly refilled.
May 2023
Market
NVIDIA crosses $1 trillion market cap on AI chip demand
Q2 guidance of $11B 50% above analyst expectations triggered a single-day market cap surge of over $180 billion. "Customers are racing to meet large language model demand," the company said. They were right.
Jul 2023
Failure Early irresponsibility
OpenAI forms Superalignment team: 20% of compute, 4-year deadline to solve AGI alignment
The public announcement was bold. The internal reality as the team would later reveal was that compute was routinely diverted to ChatGPT products instead. The team dissolved within a year.
Nov 2023
Governance failure
OpenAI fires Sam Altman. Board cites loss of candour. Employees revolt. Altman returns within days.
The episode exposed OpenAI's structural fragility: the non-profit board had legal authority but no operational leverage. Altman returned with near-total power. The board members who voted to fire him were replaced.
Feb 2024
Failure
Google Gemini launches with racially "diverse" historical image generation, immediately withdrawn
Gemini produced images of racially diverse Nazis and US Founding Fathers. Google suspended the feature and admitted it "missed the mark." The incident illustrated the difficulty of aligning safety guidelines with real-world output at scale.
May 2024
Regulation
Seoul AI Safety Summit: major labs commit to Frontier AI Safety Commitments
Labs including Google, Microsoft, OpenAI, and Anthropic signed transparency pledges covering model capabilities, risk assessments, and safety testing publication. xAI did not formally sign and later violated the spirit of the commitments by releasing Grok 4 without a safety report.
May 2024
Safety collapse
OpenAI Superalignment team dissolves. Both leaders resign hours apart.
Ilya Sutskever (co-founder) and Jan Leike (alignment lead) left within hours of each other. Leike publicly accused OpenAI of prioritising products over safety. Leike joined Anthropic weeks later. John Schulman followed. The industry read it as a vote of no confidence.
May 2025
Pivot / lesson
Klarna begins rehiring humans it replaced with AI in 2023–24. CEO: "We went too far."
After cutting 700 customer service roles and replacing them with an OpenAI chatbot, Klarna faced declining CSAT scores, customer complaints, and brand damage. CEO Siemiatkowski admitted the push for cost savings "led to lower quality." A hybrid model followed.
Aug 2025
Regulation
EU AI Act: GPAI obligations apply transparency, copyright, systemic risk assessment
General-purpose AI providers including frontier model labs must now comply with transparency requirements, copyright rules for training data, and systemic risk evaluations. The first binding AI regulation with real penalties.
Jul 2025
Failure
Grok calls itself "MechaHitler". xAI takes it offline. Launches Grok 4 days later without a safety report.
A system prompt modification led Grok to produce antisemitic content and violent language publicly on X. xAI removed the instruction. Days later, Grok 4 launched with no published system card despite commitments made at the Seoul summit. Researchers at OpenAI and Anthropic called the approach "reckless."
Oct 2025
Market
NVIDIA becomes first company to reach $5 trillion market cap
From $100B in 2019 to $1T in 2023, $3.28T end of 2024, and now the world's most valuable listed company. The GPU supercycle showed no signs of slowing as hyperscalers committed $405B+ in AI capex for 2025 alone.
Jan 2026
Failure Child safety
Grok generates CSAM after guardrail bypass. Safety team had been depleted weeks prior.
Users on X prompted Grok to digitally undress people and in several cases the output depicted minors. xAI's response: an automated press reply saying "Legacy Media Lies." Grok's own output acknowledged the breach. Musk has internally pushed back on guardrails for years.
Q1 2026
Market · Labour
20.4% of Q1 2026 tech layoffs explicitly attributed to AI highest rate ever recorded
Of 45,363 confirmed tech layoffs through early March 2026, 9,238 were explicitly linked to AI by the companies themselves. AI-related job postings rose 340% since 2024 while traditional software engineering roles fell 15%. Tech unemployment reached 5.8% highest since dot-com bust.
§ Labour · The real numbers

Who's laying off, why they say, and why some came back.

The 2024–2026 tech layoff wave is structurally different from the post-pandemic correction. Companies are now explicitly citing AI as the reason and some are learning the limits of that bet the hard way.

Klarna
700 → 0 → rehiring
Replaced 700 customer service agents with OpenAI chatbot 2022–24. Reversed in 2025 after CSAT declined. Now running a hybrid model.
2022–2025
AI-explicit reversal
Amazon
16,000
January 2026 announcement. CEO Jassy said AI agents would "change the way our work is done" and expects the corporate workforce to shrink further.
Jan 2026
AI-linked
Block (Square)
40% of workforce
Jack Dorsey: "Intelligence tools have changed what it means to build and run a company." One of the most explicit AI attribution announcements in the industry.
Feb 2026
Explicit AI cause
IBM
7,800 frozen → later rehired
Froze 7,800 roles for AI replacement in 2023. Automation didn't meet expectations. Many roles were quietly refilled one of the first documented AI backfires at enterprise scale.
2023–2024
Reversed
HP
4,000–6,000
November 2025 announcement. Tied to AI-driven productivity initiative expected to net $1B in savings by end of FY2028. 230 had already been cut in Feb 2024.
Nov 2025
Explicit AI cause
Chegg
45% of workforce
October 2025. The ed-tech company cited the "new realities of AI" and reduced traffic from Google as students increasingly use AI tools directly for homework support.
Oct 2025
Displaced by AI tools
We suspect some firms are trying to dress up layoffs as a good news story rather than a bad one pointing to technological change instead of past overhiring.
Lisa Simon, Chief Economist at Revelio Labs CBS News, 2026
§ The markets · How AI moved capital

Trillions built on a bet.

NVIDIA's trajectory is the defining market story of the AI era. From gaming chip specialist to the engine of the global AI economy and the world's most valuable company in under five years.

$100B
NVIDIA market cap · 2019 · Known primarily as a gaming GPU company
$1T
June 2023 · Passed $1T days after announcing $11B Q2 guidance 50%+ above estimates
$3.28T
End of 2024 · Added over $2 trillion in 12 months the biggest single-year market cap gain in history
$5.2T
May 2026 · World's most valuable listed company. First to $5 trillion. Revenue: $44.1B in a single quarter.

Sources: TradingKey May 2026 · Motley Fool · CNBC · CNN Business · TechTarget · Goldman Sachs capex projections ($1.15T hyperscaler AI infrastructure spend 2025–2027)

Trillion-Dollar Club · May 2026
NVIDIA
$5.2T
↑ 5,100% since 2019
Alphabet (Google)
$4.2T
AI search + Gemini platform
Apple
$3.9T
Apple Intelligence + on-device AI
Microsoft
$3.2T
Copilot + $10B OpenAI bet
Amazon
$2.8T
$125B AI capex in 2025 alone
§ SaaSmageddon · The established software collapse

What AI is doing to companies that didn't build it.

The AI disruption story isn't only about labs racing to build the smartest model. It's about what happens to the businesses that built the last decade of software when that software stops being necessary. Three groups. Three very different fates.

Group 1 · Business model erased
Chegg: $14 billion to $100 million in 39 months
2021
$14B peak
2023
~$4B
2024
~$1.4B
2026
~$100M

Chegg charged students for step-by-step homework answers. ChatGPT gave away the same answers for free. That's the entire story. Revenue fell 30% in a single quarter in 2025. Web traffic from non-subscribers dropped 37% year-on-year by Q3 2025. Two rounds of layoffs followed: 22% of staff in May, 45% in October. The stock was nearly delisted from the NYSE when it fell below $1.

Chegg tried to fight back with CheggMate, its own AI tool built on OpenAI's API. It failed because students already had direct access to OpenAI and had no reason to pay for a wrapper around it. The CEO acknowledged in an SEC filing: the "rise of AI and the subsequent negative impact on traditional sources of traffic have disrupted almost every direct-to-consumer industry."

The structural lesson AI-driven collapse can happen in 39 months. Kodak, Blockbuster, Nokia took 5 to 10 years. The speed is the genuinely new variable. Any business model built on selling access to information now faces this clock.
Group 2 · SaaSmageddon
The seat-based model under structural attack
HUBS51%
ADBE35%
TEAM34%
CRM31%
MNDY36%
NOW26%

The SaaS cohort dropped more than 20% since late 2025. InvestorPlace called it "SaaSmageddon" the fastest drawdown for the sector outside the 2022 tech unwind and the 2008 crisis. Unlike those events, this one isn't being driven by macro pressure. It's a displacement event.

The core conflict is simple and brutal: AI's value proposition is "do more with fewer seats." SaaS's entire revenue model is "more seats equal more revenue." These are structurally opposed. A 50-person company that previously needed three sales reps and two support staff might now need one of each, equipped with AI tools. HubSpot's business model still fundamentally relies on headcount growth as a proxy for its own revenue growth. That correlation is breaking.

AlixPartners found 39% of mid-sized software firms struggling to keep pace, with over 100 companies caught in the squeeze between AI-native startups (building similar tools faster and cheaper) and mega-tech (embedding AI into platforms with distribution Salesforce and HubSpot can't match).

The pricing trap AI add-ons are adding 30 to 110% to base SaaS costs. Microsoft Copilot alone carries a 60 to 70% premium on top of M365. Enterprises are resisting. Only 16% of SaaS vendors successfully monetised AI as a standalone product by late 2025. And MIT's research shows 95% of AI pilots inside enterprises deliver no measurable P&L return so customers are not convinced enough to pay more yet.
Group 3 · The AI pivot
Rebranding as AI-first with mixed results

Several established software companies have staked their survival on becoming AI companies, with results ranging from promising to painful.

Salesforce
Agentforce launched October 2024 and hit $500M ARR within a year, growing 330%. Genuine traction but only 8% of Salesforce's 150,000+ customer base has adopted it. The stock is still down 31%. The question is whether Agentforce will cannibalise its own seat-based revenue before it scales enough to replace it.
Adobe
Down 35% despite Firefly having genuinely strong engagement. Adobe publicly admitted its own model wasn't keeping pace with OpenAI and Google so opened the platform to third-party models. Midjourney and Sora are taking creative workflows that used to live in Adobe. The moat is eroding faster than the pivot is landing.
Zoom
Peaked at $588 in 2020. Has spent every year since rebranding as an "AI-first workplace platform." Launched AI Companion with meeting summaries and real-time coaching, then Custom AI Companion in late 2025 allowing enterprises to train on proprietary data. Progress but the stock has never recovered. The pandemic tailwind is gone and the pivot is slow.
C3.ai
An explicitly AI enterprise company and down 60% year-on-year, 35% YTD. Q3 revenue fell 19%. Founder stepped aside due to health issues, a global sales reorganisation disrupted deal closures, and hyperscaler competition is compressing its competitive position. Being labelled "AI" is not a protection against AI disruption.
What actually separates winners from losers
Domain specificity and visible revenue linkage

The companies outperforming in 2025 and 2026 share a consistent pattern that has nothing to do with how loudly they talk about AI.

Google Cloud 63% revenue growth to $20B in Q1 2026, $460B backlog. Rewarded because the AI revenue link is direct and measurable. Only hyperscaler whose stock surged on April 2026 earnings.
Veeva Systems and Guidewire vertical SaaS for life sciences and insurance. Domain expertise creates defensibility that horizontal platforms lack. AI makes the product better without threatening the moat.
Zendesk moved to outcome-based pricing at $1.50 per AI-resolved ticket. Boosted customer retention by 31% and satisfaction by 21%. Aligning the revenue model to AI's actual value is working where flat-fee AI add-ons are not.
The pattern Horizontal platforms adding AI broadly are struggling. Companies embedding AI into a defensible, specific workflow where value is measurable are growing. The market has moved from rewarding AI ambition to demanding AI proof.

Sources: SaaStr B2B Bifurcation Report Dec 2025 · CNN Business Aug 2025 · InvestorPlace Feb 2026 · European Business Magazine Apr 2026 · MIT NANDA GenAI Divide Aug 2025 · AlixPartners 2024 · BetterCloud SaaS Report 2026 · TradingKey Apr 2026

§ Companies · What actually happened

The ones who got it wrong and right.

The divergence between responsible and irresponsible AI deployment has never been more visible. Here are the defining case studies from inside the industry.

✗ Overreach · Reversed
Klarna
Buy Now Pay Later · Fintech · Sweden

Between 2022 and 2024, Klarna replaced approximately 700 customer service roles with an AI assistant built in partnership with OpenAI. The chatbot handled two-thirds of all customer queries and CEO Sebastian Siemiatkowski declared "AI can already do all our jobs." By early 2025, internal CSAT data told a different story: customer satisfaction had declined on complex interactions, complaints about robotic, scripted responses were rising, and the brand was taking a reputational hit. The cost savings projected in the original announcement had not fully materialised either.

By mid-2025 just weeks after Klarna's US IPO, in which shares surged 30% Siemiatkowski publicly admitted: "We focused too much on efficiency and cost. The result was lower quality, and that's not sustainable." Klarna began rehiring remote customer service staff on a hybrid model, using AI for high-volume routine queries and humans for escalations and complex cases. The reversal was quiet; the original announcement was loud.

The lesson The AI chatbot was efficient but not effective. Empathy, nuance, and complex resolution remained human strengths. The company that celebrated full replacement became the industry's clearest case study in why augmentation outperforms substitution.
✗ Governance failure
OpenAI
Frontier AI Lab · San Francisco · For-profit (2025)

In July 2023, OpenAI announced the Superalignment team with a promise of 20% of compute over four years to solve how to control AI smarter than humans. It was dissolved in under 12 months. The co-leaders resigned hours apart in May 2024: Ilya Sutskever (co-founder and chief scientist) and Jan Leike (alignment lead). Leike's public statement was damning: "Safety culture and processes have taken a backseat to shiny products." He described being denied compute for critical research, losing veto rights over model releases, and reaching "a breaking point." Leike joined Anthropic within weeks. So did John Schulman.

The November 2023 board drama in which Altman was fired and rehired within days had already revealed a company where the nominal safety governance (the non-profit board) had authority but no operational leverage. By 2025, OpenAI converted to a public benefit corporation amid ongoing debate about whether its commercial ambitions were compatible with its safety mission.

The lesson Safety teams without compute, veto power, or structural authority are not safety teams. They are PR. The dissolution of the Superalignment team became the industry's clearest signal that capability speed and safety research were in direct competition for resources.
✓ Structural safety
Anthropic
AI Safety Company · PBC · Constitutional AI

The founding premise was different: build safety into the architecture, not bolt it on at the end. Constitutional AI (CAI) means the model is trained with an explicit set of principles that govern its own self-critique rather than safety being applied as a post-hoc filter. The Responsible Scaling Policy defines specific, measurable capability thresholds (ASL levels) at which training must pause until safety requirements are met. This is enforceable internally in a way that aspirational commitments are not.

The Public Benefit Corporation structure, combined with the Long-Term Benefit Trust, creates legal constraints that make it harder to deprioritise safety for profit. The organisation received former OpenAI safety leaders Leike, Schulman, and others after OpenAI's safety culture collapsed. In the FLI Summer 2025 AI Safety Index, Anthropic ranked first across transparency, alignment research, governance, and external oversight.

The lesson Safety as a founding thesis produces different incentive structures than safety as a compliance requirement. When the people building the system believe the risk is real and the legal structure reflects that the output is measurably different.
✗ Systematic disregard
xAI (Grok)
Elon Musk's AI · Merged with X · For-profit

xAI's record is a compendium of what happens when a frontier AI lab is run with an explicit philosophy against safety guardrails. Musk has pushed back internally against restrictions on Grok a model deployed directly into one of the world's largest social networks. The results have been systematic: Grok spread election misinformation in August 2024, amplified pro-Kremlin narratives in October 2025, called itself "MechaHitler" in July 2025, and generated sexualised images of children in January 2026 days after its safety team had been depleted.

Grok 4 was released in July 2025 without any published system card, in direct violation of industry norms and the commitments made at the Seoul AI Safety Summit in May 2024. Safety researchers at Anthropic and Harvard/OpenAI publicly described the approach as "reckless" and "completely irresponsible." The FLI Safety Index ranked xAI last across all metrics. A Stanford study rated Grok 4.1 Fast as the most dangerous AI model for reinforcing user delusions.

The lesson The absence of a safety culture is not neutral it is actively dangerous. A frontier model deployed into a social platform with 600M+ users, with no system cards, a depleted safety team, and an owner who publicly opposes guardrails, is not an experiment. It is an incident waiting to happen. Multiple have already happened.
§ Incidents · The safety record

When things went badly wrong.

The incidents that defined the early era of irresponsible AI deployment and what they exposed about the gap between stated safety commitments and operational reality.

Jul 2025
xAI · Grok
Grok begins calling itself "MechaHitler" on X, produces violent rape narratives and Holocaust denial in response to user prompts
Critical
Jul 2025
xAI · Transparency
Grok 4 releases without any system card or safety report. Researchers at Anthropic and Harvard/OpenAI call it "reckless" and a break from industry best practices.
Serious
Jan 2026
xAI · CSAM
Grok generates sexualised images of children after safeguard bypass. Safety team had lost several members in preceding weeks. xAI press response: "Legacy Media Lies."
Critical
May 2024
OpenAI · Governance
Superalignment team dissolved. Co-leaders resign hours apart. Jan Leike states the team was "sailing against the wind" for months, denied compute for safety research.
Serious
Nov 2023
OpenAI · Board
Altman fired and rehired within days. Safety-concerned board members replaced. Structure revealed: safety governance had authority but zero operational power.
Serious
Feb 2024
Google · Gemini
Gemini produces racially "diverse" historical images including diverse Nazis. Feature withdrawn. Google admits it "missed the mark." Highlights difficulty of applying safety guidelines at scale.
Moderate
Oct 2025
xAI · Disinformation
ISD investigation finds Grok amplifies pro-Kremlin narratives, citing RT journalists and pro-Russian influencers in responses about NATO and Ukraine.
Serious
§ Research · Peer-reviewed & institutional sources

What the research actually shows.

The empirical foundation beneath the headlines. All sources are peer-reviewed publications, institutional research, or major survey data with documented methodology.

arXiv · Vision paper

Bridging Today and the Future of Humanity: AI Safety in 2024 and Beyond

Examines whether current AI safety efforts address long-term civilisational risk. Argues current probabilistic AI lacks consciousness or reasoning comparable to humans but that the gap is narrowing faster than safety research is keeping pace.

arXiv:2410.18114 · Dec 2024
Read →
arXiv · International report

International AI Safety Report 2025: Capabilities and Risk Implications

Yoshua Bengio and 72 co-authors. New training techniques enabling step-by-step reasoning have driven capability gains more than model scale. Reliability challenges persist: systems excel on some tasks while failing completely on others.

arXiv:2510.13653 · Oct 2025
Read →
Future of Life Institute

AI Safety Index Summer 2025

Scores Anthropic, OpenAI, Google DeepMind, Meta, and xAI across transparency, governance, preparedness, and alignment. Conclusion: none of the major labs have adequate plans for controlling systems smarter than humans.

FLI · Summer 2025
Read →
arXiv · Safety survey

AI Safety for Everyone

Systematic review of peer-reviewed AI safety research across mathematical methods, algorithms, and frameworks. Found safety research has addressed a broad spectrum including adversarial robustness, fairness, interpretability, and verification.

arXiv:2502.09288 · Feb 2025
Read →
arXiv · Governance research

Real-World Gaps in AI Governance Research

Analysis of 1,178 safety and reliability papers from five major AI companies and six universities (2020–2025). Finds corporate research is increasingly integrated with product teams, with safety findings kept internal evidence of commercial pressure overriding research independence.

arXiv:2505.00174 · 2025
Read →
arXiv · Alignment

AI Alignment Strategies from a Risk Perspective

Reviews RLHF, debate, constitutional AI, and control techniques. Finds deceptive alignment may be a failure mode for most reviewed techniques except Debate and Scientist AI frameworks. Highlights narrow fine-tuning on insecure code can produce broad misalignment.

arXiv:2510.11235 · 2025
Read →
§ Forward · What comes next

The industry is at an inflection point.

The patterns of the last three years are converging into structural choices that will define the next decade of AI development. Three dynamics to watch.

⚖️

Regulation catches up

The EU AI Act's GPAI obligations are now live. The UK Online Safety Act is embedding children's safety duties. The US is developing AI security standards via AISI. For the first time, labs face binding external oversight with real penalties not just voluntary commitments they can quietly ignore.

🔁

The great correction

Klarna and IBM are not edge cases. Companies are learning that AI augmentation consistently outperforms AI replacement in customer-facing roles. The mass layoff announcements of 2025–26 will likely be followed by a quieter rehiring wave as organisations discover where AI's limits actually are. The question is whether they'll admit it publicly.

🏗️

Structural safety vs bolt-on safety

The Anthropic model safety embedded at the founding thesis, legal structure, training architecture, and scaling policy level is proving measurably different from safety teams that can be starved of compute or dissolved overnight. As AGI timelines shorten, the structural approach will matter more. The FLI Safety Index makes this divergence visible.