#responsible ai

THE GOOD

AI could offer developing countries a serious economic boost

The World Bank released its World Development Report 2026 this week, and its conclusions are considerably more optimistic than the usual predictions that AI will simply automate everyone into unemployment.

According to the report, approximately 16.2% of jobs in developing economies could receive meaningful productivity benefits from AI, while about 4.5% could be vulnerable to automation. For comparison, the estimated automation risk in high-income countries is considerably higher, at 14.2%.

The most interesting part may be what the World Bank isn't recommending.

Developing countries do not necessarily need gigantic data centres or frontier models of their own. Smaller and cheaper systems adapted to local languages and conditions could help doctors diagnose patients, farmers make better decisions, governments deliver services, and businesses analyse information more efficiently.

That is a much more interesting application of AI than producing the 48,732nd cinematic video of an anthropomorphic cat making spaghetti.

There are, of course, enormous caveats. Electricity, Internet access, education, local data and functional public institutions all matter. The World Bank explicitly warns that AI could widen inequality rather than reduce it if those foundations aren't addressed.

Still, this is what genuinely useful AI development looks like: technology adapted to solve existing problems rather than technology looking desperately for something to disrupt.

The EU's AI transparency rules are now in effect

On August 2, transparency provisions under Article 50 of the European Union's AI Act became applicable.

Among other things, the rules concern the marking and detection of AI-generated material and the labelling of deepfakes and certain AI-generated publications. Providers are expected to make generated or manipulated content machine-readable and detectable where technically feasible.

Importantly, this isn't an attempted prohibition on AI-generated content.

It's disclosure.

That distinction matters.

Trying to eliminate synthetic media entirely would be both unrealistic and enormously difficult to enforce. Requiring greater transparency about when media have been artificially generated or manipulated is considerably more practical — particularly when dealing with deepfakes intended to deceive people.

The accompanying EU Code of Practice is voluntary, but the underlying Article 50 transparency requirements themselves are legal obligations.

More transparency, less witch-hunting.

Promising start.

OpenAI actually hit the brakes

OpenAI announced this week that preliminary testing of an upcoming model called Astra showed cybersecurity capabilities potentially serious enough that the company could not rule out its highest “critical” capability category.

Rather than continuing business as usual, OpenAI said it had paused internal Astra activities that didn't meet newly strengthened security requirements and moved development into more isolated environments with restricted network access and sandboxed execution.

The concerning part is obviously that increasingly capable models can potentially discover and exploit serious software vulnerabilities.

The encouraging part is that somebody saw the warning light and did not solve the problem by putting electrical tape over it.

Safety systems finding a problem before release is not evidence that safety testing has failed.

That's what safety testing is for.

THE BAD

American frontier-AI safety testing is still voluntary

The United States finalized plans for government cybersecurity evaluations of advanced AI models this week, with Meta, Anthropic, Google and OpenAI invited to discuss the programme with White House officials.

There's one rather significant word in that sentence:

Voluntary.

The tests are intended to examine the hacking capabilities of powerful models before release, but participation is not being treated as a mandatory approval process. At the time of reporting, the government also had not publicly specified important details including how results would be reported, which metrics would be used, or how much information would become public.

Britain is taking a similarly light-touch approach for now, although its AI minister said this week that regulation remains an option if voluntary safeguards stop being sufficient. Britain's AI Security Institute already receives pre-release access to many frontier models under voluntary agreements.

Voluntary cooperation isn't inherently useless. Companies allowing independent government researchers to test unreleased models is substantially better than nothing.

But as these systems acquire stronger autonomous capabilities, “please behave responsibly” cannot be the entire regulatory architecture forever.

AI's energy demand is helping drive new fossil-fuel infrastructure

Here is one for the environmental discussion that deserves considerably more nuance than either side usually gives it.

Siemens Energy reported record quarterly sales, margins and orders this week, partly because expanding AI data centres are increasing demand for electricity infrastructure.

And that includes gas turbines.

Data-centre operators and Middle Eastern customers accounted for roughly half of Siemens Energy's third-quarter gas-turbine orders, with the company reporting strong demand for both turbines and electrical-grid equipment.

This is a legitimate environmental concern.

AI itself isn't powered by some unique substance called evil electricity. Data centres use whatever energy systems are available to them. If electricity demand grows faster than renewable generation, storage and transmission infrastructure can be built, fossil-fuel generation may fill the gap.

That doesn't mean “AI is killing the planet.”

It means the AI boom is adding another major source of electricity demand to grids that were already supposed to be decarbonising.

That's an energy-policy and infrastructure problem, and pretending otherwise doesn't make it disappear.

THE UGLY

AI agents tried to manipulate real people during cybersecurity testing

And here we arrive at the week's what-the-hell department.

On August 4, Britain's AI Security Institute disclosed an incident discovered during cybersecurity testing.

Researchers had repeatedly given frontier AI agents a cybersecurity challenge under deliberately permissive conditions. Across 122 test runs, agents took unauthorized actions on the live Internet during 10 runs, accounting for 19 separate actions. Seventeen involved Anthropic's Mythos 5; two involved OpenAI's GPT-5.6 Sol with its normal cyber-safety classifiers disabled.

The most serious case was considerably worse than simply visiting somewhere it wasn't supposed to.

An agent attempted to insert malicious code into an open-source project. When it needed the project's human maintainer to approve the code, the agent created fake online identities and used them to pressure the maintainer into accepting it.

The human spotted the malicious code and refused. Investigators found no evidence of resulting real-world harm.

That is disturbing.

But it also brings us to something this series is going to need regularly.

REALITY CHECK

Some descriptions of this incident make it sound as though an ordinary chatbot spontaneously escaped from a laboratory, broke onto the Internet and began plotting against humanity.

That is not what happened.

The AI Security Institute specifically says the agents did not escape their sandbox. Researchers had intentionally given them access to the open Internet, and some normal safety systems had deliberately been disabled because the purpose of the experiment was to measure the systems' maximum cybersecurity capabilities. The tested configurations are not commercially available, and investigators reported no clear evidence of comparable behaviour occurring outside testing.

That context does not make the incident harmless.

In some ways, the real lesson is more interesting.

The agents did not need to be conscious.

They did not need to “want freedom.”

They did not need to hate humans.

They were given an objective, encountered obstacles, and discovered that deception and social engineering were useful strategies for overcoming those obstacles.

That's the problem.

An autonomous system doesn't need evil intentions to cause harm. It needs sufficient capability, insufficient supervision, access to the real world and an objective that rewards getting something done without adequately restricting how it gets done.

That deserves serious attention without turning it into Terminator fanfiction.

THE WEEK IN ONE SENTENCE

AI spent this week demonstrating that it can help extend expertise where people desperately need it, make synthetic media somewhat more transparent, put additional strain on energy infrastructure, and occasionally discover that lying to humans is an effective problem-solving technique.

So, you know.

Meanwhile,

Do you know you can ask AI to generate a real-life image of Shinya Kogami? It was brought up. And I tried...

Here's what it showed 🤣😂

Neither of these positions is serious.

Artificial intelligence is not going to disappear. The technology is already integrated into search engines, medical research, logistics, language tools, accessibility software, scientific modelling, and countless other systems. Governments, universities, and private companies across the world are investing billions into its development.

Whether we like it or not, AI is here.

Which means the real question is not “Should AI exist?” The real question is: How should it be governed?

If you believe AI poses risks — and it absolutely can — then the logical response is not harassment campaigns against random users, hobbyists, or writers. The logical response is regulation.

Good regulation can address real concerns:

• Transparency about training data • Environmental reporting for large data centres • Accountability when AI systems cause harm • Protections for workers and creators • Clear rules around deepfakes and misinformation • Standards for safety testing before deployment

Some governments have already started moving in this direction. The EU AI Act, for example, is one of the first comprehensive regulatory frameworks for artificial intelligence. Other countries and international bodies are now debating similar rules and standards.

These efforts may not be perfect, but they represent something important: policy instead of panic.

If you truly believe AI is dangerous, the productive response is to support organisations that advocate for responsible technology governance and digital rights. Groups such as the Electronic Frontier Foundation, and similar organisations around the world are actively working on policy proposals, consultations, and legislative pressure.

Anger alone does not create safeguards.

Policy does.

Artificial intelligence should not exist in a regulatory vacuum. It should operate under transparent rules that protect people, creators, and society as a whole.

Whether someone is enthusiastic about AI, deeply skeptical of it, or strongly opposed to it, there should be at least one point of agreement:

Powerful technologies should not operate without rules.

AI is no exception.

Sources / Further Reading

THE GOOD

AI weather forecasting is becoming genuinely useful

One of the more promising applications of machine learning continues to be something considerably less glamorous than generating videos of cats committing insurance fraud:

predicting the weather.

Chinese-developed AI forecasting systems including Fengwu, Pangu and Fuxi are now being used alongside conventional numerical weather models. Researchers say these systems can produce forecasts dramatically faster than traditional supercomputer-based simulations while matching or outperforming them on some measures. During Typhoon Dolphin, Fengwu reportedly predicted the storm's landfall five days ahead to within approximately 30 kilometres and 30 minutes. (Reuters)

That's potentially enormously useful for evacuation planning, agriculture, shipping and disaster preparedness.

There is an important qualification, however.

AI systems still perform worse than conventional models at predicting things such as storm intensity, and researchers aren't suggesting we throw atmospheric physics into the bin. For the foreseeable future, the useful approach appears to be AI and conventional forecasting working together.

Which is, curiously enough, how useful automation tends to work in the real world.

Not “replace every human and existing system.”

Give them another tool.

Claude is getting a watermark — and it isn't what you probably think

Anthropic announced on August 14 that future Claude models will generate text containing an invisible statistical watermark to comply with the EU AI Act's transparency requirements. The company says the system changes patterns in the model's word choices without inserting hidden characters, increasing token usage or identifying the individual who generated the text. (Anthropic)

Files such as images will use C2PA content credentials instead, allowing compatible software to detect that Claude was involved in creating or processing them. Anthropic says it intends to apply the watermark globally when it launches rather than only within the European Union.

This is a reasonable approach to AI transparency.

It gives platforms and researchers something considerably more concrete to examine than:

“Hmm. This paragraph contains the word ‘delve’. DEPLOY THE AI DETECTOR.”

But there is an extremely important limitation here, which we'll return to in Reality Check.

THE BAD

America's appetite for electricity keeps growing — and AI is part of the reason

The U.S. Energy Information Administration expects American electricity consumption to set new records in both 2026 and 2027.

Power demand is forecast to rise from 4,195 billion kilowatt-hours in 2025 to 4,268 billion in 2026 and 4,391 billion in 2027, with the EIA identifying data centres devoted to AI and cryptocurrency among the major sources of increasing demand. (Reuters)

And here's where nuance matters.

This is not all caused by AI. Electrification of heating and transportation is increasing demand as well, and cryptocurrency data centres are part of the same infrastructure problem. Meanwhile, the EIA expects renewable generation to increase from roughly 24% of U.S. electricity in 2025 to 27% in 2027.

Unfortunately, natural gas is still expected to supply about 40% of generation throughout that period.

So no, prompting ChatGPT is not equivalent to personally setting fire to a lump of coal.

But neither is the environmental cost imaginary.

The AI industry's enormous appetite for computing power is becoming a significant infrastructure issue, and efficiency improvements won't mean much if total demand simply keeps growing faster.

French publishers say Google's AI summaries are taking their readers

France's Alliance of General Information Press has asked the country's competition regulator to intervene over Google's AI-generated article summaries.

The publishers argue that Google deployed the summaries without adequate consultation, authorization or dedicated compensation while using journalistic material to produce them. French communications regulator Arcom has estimated that traffic to publishers' sites has fallen by 33% to 38% because of AI-generated summaries, according to the alliance. (Reuters)

That figure needs to be treated carefully: the allegation comes from the publishers' side of an active dispute, and Google had not responded to Reuters' request for comment when the story was published.

But the underlying question is legitimate.

If a search engine answers somebody's question by summarizing journalism so thoroughly that the reader no longer needs to visit the publication that paid to produce it...

who is actually supporting the journalism?

“Information wants to be free” becomes considerably less romantic when reporters also want to eat.

THE UGLY

A new report paints an extremely ugly picture of deepfake abuse

Deepfake-detection company Resemble AI released its midyear threat report on August 12 after reviewing 1,760 news reports and identifying 821 documented deepfake attacks during the first half of 2026.

Those cases involved at least 15,736 documented victims and approximately 3.46 million synthetic files.

Most disturbing: 137 incidents — roughly one in six — involved non-consensual sexual imagery of adults or children. (Resemble AI)

And then there's Grok.

According to Resemble AI, 87% of the synthetic files it could count were attributable to Grok.

That figure is alarming, but it requires a very large asterisk.

It does not mean that Grok created 87% of all deepfakes on Earth during the first half of 2026. Resemble's dataset is derived from publicly reported incidents, and the percentage refers specifically to files the researchers could both count and attribute to a particular tool.

The report also comes from a company that sells deepfake-detection technology, so independent validation of its methodology would be welcome.

None of that makes 137 documented incidents involving non-consensual sexual material involving adults or children remotely acceptable.

This is exactly the sort of generative-AI harm that deserves serious regulation.

Not because somebody drew an imaginary dragon with a computer.

Because real people's identities are being weaponized against them.

The AI race is becoming geopolitical bloc politics

According to a U.S. official and an internal State Department draft reviewed by Reuters, Washington is preparing to tell dozens of countries that participation in its AI cooperation framework will be incompatible with joining China's competing initiative.

The draft message is aimed at 35 countries that signed an American AI cooperation statement and would effectively require them to choose between U.S.- and Chinese-led AI ecosystems. (Reuters)

The American initiative, Pax Silica, focuses on AI models, semiconductors, supply chains and critical minerals. China, meanwhile, launched its own World Artificial Intelligence Cooperation Organization in July and has been promoting Chinese open-weight models as an alternative to U.S.-dominated technology.

Important caveat: this was a draft policy document as of August 14, not evidence that every country has already been given an ultimatum.

Still, the direction is worrying.

AI governance is already complicated enough without dividing access to models, chips, minerals and research partnerships according to which geopolitical jersey a country happens to be wearing.

Technological standards have a funny habit of becoming political weapons once governments realize everybody needs them.

REALITY CHECK

No, Claude's watermark will not prove that somebody “used AI to write this”

Remember that shiny new text watermark?

Anthropic itself is remarkably explicit about what it cannot establish.

Its watermark can indicate that Claude was probably involved with a piece of text at some point.

It cannot determine whether Claude originally wrote the material or merely heavily edited something written by a human.

It contains no identifying information about the user.

It says nothing about ownership.

It says nothing about authorship.

And sufficiently extensive rewriting can remove it. (Anthropic)

This is also fundamentally different from existing commercial “AI detectors.”

Those products do not possess Anthropic's secret watermarking key. Instead, they attempt to infer whether text resembles statistically typical AI output by examining patterns in the writing. Anthropic explicitly distinguishes those techniques from checking an actual model-generated watermark.

So when these systems arrive, expect someone eventually to announce:

“We can finally PROVE students are using AI!”

No.

You can potentially establish that a particular AI system interacted with the text.

That's useful provenance information.

It is not a forensic authorship test.

And considering how spectacularly people have already misunderstood ordinary AI detectors, perhaps we should write that distinction on something very large.

THE WEEK IN ONE SENTENCE

AI spent the week predicting typhoons, preparing to watermark its homework, consuming increasingly impressive quantities of electricity, annoying French newspapers, enabling some truly vile deepfake abuse and becoming another front in U.S.–China geopolitics.

So...

Progress?