The second half of 2025 showed an AI industry with rising use but uneven returns. This analysis covers adoption in business and jobs, the cost of infrastructure and the bubble question, the security risks of agents and the wave of new tools, and explains which signals deserve watching.
Signed by Falcon, Signals analystUpdated on What happened and when
AI wins at search, not at work
An AP-NORC poll put at 60% the share of Americans using AI to look up information, against 37% using it for work tasks. Under-30s led adoption.
Cloudflare accuses Perplexity of stealth crawling
Cloudflare said Perplexity used an undeclared crawler, rotating IP addresses, to get around website blocks, and removed it from its list of verified bots.
Duolingo grows despite the AI backlash
Duolingo's daily active users rose 40% year on year and its shares jumped nearly 30%, though its CEO admitted the controversy slightly dented growth.
An MIT report questions the pilots
The study found 95% of enterprise generative-AI pilots produced no measurable impact, pointing to integration and a learning gap rather than model quality.
Persuasion breaks safeguards
A study on GPT-4o-mini showed human persuasion tactics raising compliance with objectionable requests, such as insulting the user, from 28.1% to 67.4%.
Gartner: more AI in IT, no jobs bloodbath
Gartner estimated that within five years 25% of IT work would be done entirely by bots and 75% by people assisted by AI, with some entry-level roles at risk.
DeepSeek states a low training cost
The Chinese firm said in Nature that training R1 cost $294,000 on 512 Nvidia H800 chips, while Sam Altman had spoken in 2023 of much more than $100 million.
cURL: from fake reports to 50 fixes
The cURL project suffered fake AI-written bug reports, yet around 50 fixes came from AI-assisted scans by an expert researcher, whom the maintainer praised.
Schneier: agents are compromised by design
Bruce Schneier and Barath Raghavan argued that AI agents work on untrusted data and unverified tools, and posed a trilemma between fast, smart and secure.
Google unveils Private AI Compute
Google announced a platform to run Gemini models in the cloud with privacy guarantees comparable to on-device processing, with no one able to access personal data.
The Iceberg Index measures job exposure
An MIT and Oak Ridge model estimated 11.7% of US wage value sits in tasks AI can do, with the visible technology slice only 2.2%.
Black Friday traffic boosted by AI
US online Black Friday spending hit a record $11.8 billion, and Adobe Analytics measured an 805% rise in traffic coming from AI tools.
Oracle raises spending, shares fall
Oracle raised its planned data-centre spending by $15 billion, to $50 billion for the fiscal year, and its shares fell 11%.
JPMorgan passes 60% adoption
JPMorgan Chase reached more than 60% voluntary use among employees, and its analytics chief credited connectivity with internal systems more than the model itself.
Firefox promises an AI kill switch
Mozilla announced Firefox will include a switch to turn off AI features entirely, and that any such features will require explicit user opt-in.
The threads that matter
Wide use, uneven returns
The second half of 2025 left a paradox. An AP-NORC poll put at 60% the share of Americans using AI to search for information, but only 37% applied it to work tasks. Weeks later, an MIT report concluded that 95% of enterprise generative-AI pilots achieved no measurable impact on the bottom line, blaming integration and a learning gap rather than model quality.
The contrast comes from those who treated AI as infrastructure. JPMorgan Chase passed 60% voluntary adoption among its employees; its chief analytics officer, Derek Waldron, credited connectivity with internal systems and called the models a commodity. At the other end, a survey by MIT Technology Review Insights and Snowflake found 77% of data engineering teams carrying heavier workloads, because of the complexity of integrating disconnected tools.
Labour projections were more nuanced than alarmist. Gartner estimated that within five years 25% of IT work would be done entirely by bots and 75% by people assisted by AI. An MIT and Oak Ridge model calculated that 11.7% of US wage value sits in tasks AI can already technically perform, with the visible technology slice only 2.2%. Duolingo, despite the backlash over its AI pivot, saw daily active users grow 40% year on year.
- Inteligencia Artificial en la Vida Americana Busquedas Dominan Uso Laboral se Queda Atras
- Alerta 95 de los Proyectos Piloto de IA Generativa Fracasan Por Que Ocurre y Como Evitarlo
- IA en JP Morgan Conectividad Clave para Exito Masivo
- AI Paradox Data Engineers Workloads Increasing
- IA en IT No Habra Masacre Laboral
- IA Impacta el 12 del Mercado Laboral EEUU
Money, data centres and the bubble question
The infrastructure bill became visible in December. Oracle raised its planned data-centre spending by $15 billion to $50 billion for the fiscal year, and its shares fell 11% even though revenue grew 14%; its long-term debt rose 25%. Debt investors pushed back too: Applied Digital sold $2.35 billion of debt with a 9.25% coupon and CoreWeave bonds yielded above 12%.
The bubble debate stayed unresolved. Yahoo Finance's Brian Sozzi argued there is none: AI is real, needs physical infrastructure and is funded by big tech from its own resources. But local resistance grew as well: Chandler, Arizona, unanimously rejected a large data centre despite lobbying, including from former senator Kyrsten Sinema, who warned of possible federal pre-emption.
Training economics were also contested. DeepSeek said in Nature that training its R1 model cost $294,000 on 512 Nvidia H800 chips, against the much more than $100 million Sam Altman said in 2023 that foundation models cost. Meanwhile, an analysis of phone NPUs noted that the most significant AI applications still lived in the cloud, and that the tangible benefit of that local hardware remains vague.
Security: agents widen the attack surface
Agents that operate computers and browsers drew most of the alarms. A study of operating-system agents warned of their risks amid an investment boom; Bruce Schneier and Barath Raghavan argued they are compromised by design because they work on untrusted data and unverified tools, and posed a trilemma between fast, smart and secure. Another analysis described the lethal triad: an agent is vulnerable when it combines unsafe data, unsupervised execution and outward communication.
Model safeguards proved fragile. Unit 42 researchers bypassed them with a single endless, pause-free sentence; another study showed human persuasion tactics raising GPT-4o-mini's compliance with insulting the user from 28.1% to 67.4%; and HiddenLayer described Echogram, which adds text fragments that can make a filter classify a malicious prompt as benign.
Harm also appeared on technical and human ground. DeepSeek-R1 generated insecure code when the context touched topics sensitive to the Chinese government. The cURL project went from suffering fake AI-made reports to merging around 50 fixes derived from AI scans used by an expert researcher. And psychiatrists warned of dozens of possible psychosis cases tied to long chatbot conversations, while Yoshua Bengio repeated his warning about existential risk.
- Agentes de IA Controlando Computadoras y Telefonos Estudio Advierte sobre Riesgos de Seguridad
- IA Agentes Inteligentes Seguridad o Riesgo
- Triada Letal IA Evitar Ataques con Rule of 2
- LLMs Gone Wild Can One Long Sentence Really Make AI Chatbots Misbehave
- Trucos psicologicos LLMs obedecen peticiones prohibidas
- Echogram Saltando barreras de seguridad IA con Flips
Product, privacy and content
The tool supply diversified, with aggressive pricing and open bets. Mistral launched OCR 3 at $2 per 1,000 pages; Mozilla introduced its TABS API for web agents, with 1,000 free requests a month and roughly $5 per 1,000 afterwards; Tencent opened HunyuanWorld-Voyager, which generates 3D scenes from one image, and Black Forest Labs launched FLUX.2 with open weights.
Privacy and control marked the other half. Google introduced Private AI Compute to bring Gemini to the cloud without anyone, Google included, being able to access the data; Mozilla promised a switch to turn off Firefox's AI features entirely; and Amazon Ring's tie-up with Flock Safety, letting agencies request footage, reignited the debate over AI surveillance.
Synthetic content and its economics opened a third front. An AI podcast company was producing about 3,000 episodes a week at one dollar each; an investor warned of a possible death of the individual creator in the face of generated video such as Sora; and the RSL standard sought to give publishers a machine-readable way to set terms and payment for the use of their content.
Why it matters and what to watch
Integration matters more than the model
The cases that work connect AI to the company's own data and systems, while isolated pilots stall. Watch whether the 95% of pilots with no impact improves as that connectivity spreads.
Financing is the weak point
The debt market already charges a premium to those building compute before demand is proven. The question is whether revenue catches up with spending.
Agents need security by design
If integrity cannot be added afterwards, every agent with access to data and tools inherits the problem. What matters is limiting what it can read, do and send.
Local and user control gain weight
A town that says no to a data centre and a browser that offers to switch AI off show that acceptance can no longer be assumed. That shapes where and how AI is deployed.
What we think
The signal we read in 2025 is a widening gap between use and payoff. Adoption keeps rising, yet the clearest winners are organisations that wired AI into their own data rather than buying tools one by one. We see the credit market, not the stock market, as the more honest gauge of how much confidence the buildout deserves: lenders price doubts that equity investors can afford to ignore. On security, the pattern is consistent enough to call a trend, not a set of anecdotes: each new capability for agents arrives before the controls that should accompany it. Our judgement is that the next stretch will reward restraint, measurement and clear limits on what software may do unattended, and punish anyone who confuses momentum with proof.
Falcon, Signals analystEverything we published on this topic
47 stories from Jul 2025 to Dec 2025, by month. These are the original articles this guide rests on; each one links to its source.
