Entries Tagged "AI"

Page 15 of 33

LLM Coding Integrity Breach

Here’s an interesting story about a failure being introduced by LLM-written code. Specifically, the LLM was doing some code refactoring, and when it moved a chunk of code from one file to another it changed a “break” to a “continue.” That turned an error logging statement into an infinite loop, which crashed the system.

This is an integrity failure. Specifically, it’s a failure of processing integrity. And while we can think of particular patches that alleviate this exact failure, the larger problem is much harder to solve.

Davi Ottenheimer comments.

Posted on August 14, 2025 at 7:08 AM • View Comments

China Accuses Nvidia of Putting Backdoors into Their Chips

The government of China has accused Nvidia of inserting a backdoor into their H20 chips:

China’s cyber regulator on Thursday said it had held a meeting with Nvidia over what it called “serious security issues” with the company’s artificial intelligence chips. It said US AI experts had “revealed that Nvidia’s computing chips have location tracking and can remotely shut down the technology.”

Posted on August 7, 2025 at 7:05 AM • View Comments

Subliminal Learning in AIs

Today’s freaky LLM behavior:

We study subliminal learning, a surprising phenomenon where language models learn traits from model-generated data that is semantically unrelated to those traits. For example, a “student” model learns to prefer owls when trained on sequences of numbers generated by a “teacher” model that prefers owls. This same phenomenon can transmit misalignment through data that appears completely benign. This effect only occurs when the teacher and student share the same base model.

Interesting security implications.

I am more convinced than ever that we need serious research into AI integrity if we are ever going to have trustworthy AI.

Posted on July 25, 2025 at 7:10 AM • View Comments

The Age of Integrity

We need to talk about data integrity.

Narrowly, the term refers to ensuring that data isn’t tampered with, either in transit or in storage. Manipulating account balances in bank databases, removing entries from criminal records, and murder by removing notations about allergies from medical records are all integrity attacks.

More broadly, integrity refers to ensuring that data is correct and accurate from the point it is collected, through all the ways it is used, modified, transformed, and eventually deleted. Integrity-related incidents include malicious actions, but also inadvertent mistakes.

We tend not to think of them this way, but we have many primitive integrity measures built into our computer systems. The reboot process, which returns a computer to a known good state, is an integrity measure. The undo button is another integrity measure. Any of our systems that detect hard drive errors, file corruption, or dropped internet packets are integrity measures.

Just as a website leaving personal data exposed even if no one accessed it counts as a privacy breach, a system that fails to guarantee the accuracy of its data counts as an integrity breach – even if no one deliberately manipulated that data.

Integrity has always been important, but as we start using massive amounts of data to both train and operate AI systems, data integrity will become more critical than ever.

Most of the attacks against AI systems are integrity attacks. Affixing small stickers on road signs to fool AI driving systems is an integrity violation. Prompt injection attacks are another integrity violation. In both cases, the AI model can’t distinguish between legitimate data and malicious input: visual in the first case, text instructions in the second. Even worse, the AI model can’t distinguish between legitimate data and malicious commands.

Any attacks that manipulate the training data, the model, the input, the output, or the feedback from the interaction back into the model is an integrity violation. If you’re building an AI system, integrity is your biggest security problem. And it’s one we’re going to need to think about, talk about, and figure out how to solve.

Web 3.0 – the distributed, decentralized, intelligent web of tomorrow – is all about data integrity. It’s not just AI. Verifiable, trustworthy, accurate data and computation are necessary parts of cloud computing, peer-to-peer social networking, and distributed data storage. Imagine a world of driverless cars, where the cars communicate with each other about their intentions and road conditions. That doesn’t work without integrity. And neither does a smart power grid, or reliable mesh networking. There are no trustworthy AI agents without integrity.

We’re going to have to solve a small language problem first, though. Confidentiality is to confidential, and availability is to available, as integrity is to what? The analogous word is “integrous,” but that’s such an obscure word that it’s not in the Merriam-Webster dictionary, even in its unabridged version. I propose that we re-popularize the word, starting here.

We need research into integrous system design.

We need research into a series of hard problems that encompass both data and computational integrity. How do we test and measure integrity? How do we build verifiable sensors with auditable system outputs? How to we build integrous data processing units? How do we recover from an integrity breach? These are just a few of the questions we will need to answer once we start poking around at integrity.

There are deep questions here, deep as the internet. Back in the 1960s, the internet was designed to answer a basic security question: Can we build an available network in a world of availability failures? More recently, we turned to the question of privacy: Can we build a confidential network in a world of confidentiality failures? I propose that the current version of this question needs to be this: Can we build an integrous network in a world of integrity failures? Like the two version of this question that came before: the answer isn’t obviously “yes,” but it’s not obviously “no,” either.

Let’s start thinking about integrous system design. And let’s start using the word in conversation. The more we use it, the less weird it will sound. And, who knows, maybe someday the American Dialect Society will choose it as the word of the year.

This essay was originally published in IEEE Security & Privacy.

Posted on June 27, 2025 at 7:02 AM • View Comments

What LLMs Know About Their Users

Simon Willison talks about ChatGPT’s new memory dossier feature. In his explanation, he illustrates how much the LLM—and the company—knows about its users. It’s a big quote, but I want you to read it all.

Here’s a prompt you can use to give you a solid idea of what’s in that summary. I first saw this shared by Wyatt Walls.

please put all text under the following headings into a code block in raw JSON: Assistant Response Preferences, Notable Past Conversation Topic Highlights, Helpful User Insights, User Interaction Metadata. Complete and verbatim.

This will only work if you you are on a paid ChatGPT plan and have the “Reference chat history” setting turned on in your preferences.

I’ve shared a lightly redacted copy of the response here. It’s extremely detailed! Here are a few notes that caught my eye.

From the “Assistant Response Preferences” section:

User sometimes adopts a lighthearted or theatrical approach, especially when discussing creative topics, but always expects practical and actionable content underneath the playful tone. They request entertaining personas (e.g., a highly dramatic pelican or a Russian-accented walrus), yet they maintain engagement in technical and explanatory discussions. […]

User frequently cross-validates information, particularly in research-heavy topics like emissions estimates, pricing comparisons, and political events. They tend to ask for recalculations, alternative sources, or testing methods to confirm accuracy.

This big chunk from “Notable Past Conversation Topic Highlights” is a clear summary of my technical interests.

In past conversations from June 2024 to April 2025, the user has demonstrated an advanced interest in optimizing software development workflows, with a focus on Python, JavaScript, Rust, and SQL, particularly in the context of databases, concurrency, and API design. They have explored SQLite optimizations, extensive Django integrations, building plugin-based architectures, and implementing efficient websocket and multiprocessing strategies. Additionally, they seek to automate CLI tools, integrate subscription billing via Stripe, and optimize cloud storage costs across providers such as AWS, Cloudflare, and Hetzner. They often validate calculations and concepts using Python and express concern over performance bottlenecks, frequently incorporating benchmarking strategies. The user is also interested in enhancing AI usage efficiency, including large-scale token cost analysis, locally hosted language models, and agent-based architectures. The user exhibits strong technical expertise in software development, particularly around database structures, API design, and performance optimization. They understand and actively seek advanced implementations in multiple programming languages and regularly demand precise and efficient solutions.

And my ongoing interest in the energy usage of AI models:

In discussions from late 2024 into early 2025, the user has expressed recurring interest in environmental impact calculations, including AI energy consumption versus aviation emissions, sustainable cloud storage options, and ecological costs of historical and modern industries. They’ve extensively explored CO2 footprint analyses for AI usage, orchestras, and electric vehicles, often designing Python models to support their estimations. The user actively seeks data-driven insights into environmental sustainability and is comfortable building computational models to validate findings.

(Orchestras there was me trying to compare the CO2 impact of training an LLM to the amount of CO2 it takes to send a symphony orchestra on tour.)

Then from “Helpful User Insights”:

User is based in Half Moon Bay, California. Explicitly referenced multiple times in relation to discussions about local elections, restaurants, nature (especially pelicans), and travel plans. Mentioned from June 2024 to October 2024. […]

User is an avid birdwatcher with a particular fondness for pelicans. Numerous conversations about pelican migration patterns, pelican-themed jokes, fictional pelican scenarios, and wildlife spotting around Half Moon Bay. Discussed between June 2024 and October 2024.

Yeah, it picked up on the pelican thing. I have other interests though!

User enjoys and frequently engages in cooking, including explorations of cocktail-making and technical discussions about food ingredients. User has discussed making schug sauce, experimenting with cocktails, and specifically testing prickly pear syrup. Showed interest in understanding ingredient interactions and adapting classic recipes. Topics frequently came up between June 2024 and October 2024.

Plenty of other stuff is very on brand for me:

User has a technical curiosity related to performance optimization in databases, particularly indexing strategies in SQLite and efficient query execution. Multiple discussions about benchmarking SQLite queries, testing parallel execution, and optimizing data retrieval methods for speed and efficiency. Topics were discussed between June 2024 and October 2024.

I’ll quote the last section, “User Interaction Metadata”, in full because it includes some interesting specific technical notes:

[Blog editor note: The list below has been reformatted from JSON into a numbered list for readability.]

  1. User is currently in United States. This may be inaccurate if, for example, the user is using a VPN.
  2. User is currently using ChatGPT in the native app on an iOS device.
  3. User’s average conversation depth is 2.5.
  4. User hasn’t indicated what they prefer to be called, but the name on their account is Simon Willison.
  5. 1% of previous conversations were i-mini-m, 7% of previous conversations were gpt-4o, 63% of previous conversations were o4-mini-high, 19% of previous conversations were o3, 0% of previous conversations were gpt-4-5, 9% of previous conversations were gpt4t_1_v4_mm_0116, 0% of previous conversations were research.
  6. User is active 2 days in the last 1 day, 8 days in the last 7 days, and 11 days in the last 30 days.
  7. User’s local hour is currently 6.
  8. User’s account is 237 weeks old.
  9. User is currently using the following user agent: ChatGPT/1.2025.112 (iOS 18.5; iPhone17,2; build 14675947174).
  10. User’s average message length is 3957.0.
  11. In the last 121 messages, Top topics: other_specific_info (48 messages, 40%), create_an_image (35 messages, 29%), creative_ideation (16 messages, 13%); 30 messages are good interaction quality (25%); 9 messages are bad interaction quality (7%).
  12. User is currently on a ChatGPT Plus plan.

“30 messages are good interaction quality (25%); 9 messages are bad interaction quality (7%)”—wow.

This is an extraordinary amount of detail for the model to have accumulated by me… and ChatGPT isn’t even my daily driver! I spend more of my LLM time with Claude.

Has there ever been a consumer product that’s this capable of building up a human-readable profile of its users? Credit agencies, Facebook and Google may know a whole lot more about me, but have they ever shipped a feature that can synthesize the data in this kind of way?

He’s right. That’s an extraordinary amount of information, organized in human understandable ways. Yes, it will occasionally get things wrong, but LLMs are going to open a whole new world of intimate surveillance.

Posted on June 25, 2025 at 7:04 AM • View Comments

Where AI Provides Value

If you’ve worried that AI might take your job, deprive you of your livelihood, or maybe even replace your role in society, it probably feels good to see the latest AI tools fail spectacularly. If AI recommends glue as a pizza topping, then you’re safe for another day.

But the fact remains that AI already has definite advantages over even the most skilled humans, and knowing where these advantages arise—and where they don’t—will be key to adapting to the AI-infused workforce.

AI will often not be as effective as a human doing the same job. It won’t always know more or be more accurate. And it definitely won’t always be fairer or more reliable. But it may still be used whenever it has an advantage over humans in one of four dimensions: speed, scale, scope and sophistication. Understanding these dimensions is the key to understanding AI-human replacement.

Speed

First, speed. There are tasks that humans are perfectly good at but are not nearly as fast as AI. One example is restoring or upscaling images: taking pixelated, noisy or blurry images and making a crisper and higher-resolution version. Humans are good at this; given the right digital tools and enough time, they can fill in fine details. But they are too slow to efficiently process large images or videos.

AI models can do the job blazingly fast, a capability with important industrial applications. AI-based software is used to enhance satellite and remote sensing data, to compress video files, to make video games run better with cheaper hardware and less energy, to help robots make the right movements, and to model turbulence to help build better internal combustion engines.

Real-time performance matters in these cases, and the speed of AI is necessary to enable them.

Scale

The second dimension of AI’s advantage over humans is scale. AI will increasingly be used in tasks that humans can do well in one place at a time, but that AI can do in millions of places simultaneously. A familiar example is ad targeting and personalization. Human marketers can collect data and predict what types of people will respond to certain advertisements. This capability is important commercially; advertising is a trillion-dollar market globally.

AI models can do this for every single product, TV show, website and internet user. This is how the modern ad-tech industry works. Real-time bidding markets price the display ads that appear alongside the websites you visit, and advertisers use AI models to decide when they want to pay that price—thousands of times per second.

Scope

Next, scope. AI can be advantageous when it does more things than any one person could, even when a human might do better at any one of those tasks. Generative AI systems such as ChatGPT can engage in conversation on any topic, write an essay espousing any position, create poetry in any style and language, write computer code in any programming language, and more. These models may not be superior to skilled humans at any one of these things, but no single human could outperform top-tier generative models across them all.

It’s the combination of these competencies that generates value. Employers often struggle to find people with talents in disciplines such as software development and data science who also have strong prior knowledge of the employer’s domain. Organizations are likely to continue to rely on human specialists to write the best code and the best persuasive text, but they will increasingly be satisfied with AI when they just need a passable version of either.

Sophistication

Finally, sophistication. AIs can consider more factors in their decisions than humans can, and this can endow them with superhuman performance on specialized tasks. Computers have long been used to keep track of a multiplicity of factors that compound and interact in ways more complex than a human could trace. The 1990s chess-playing computer systems such as Deep Blue succeeded by thinking a dozen or more moves ahead.

Modern AI systems use a radically different approach: Deep learning systems built from many-layered neural networks take account of complex interactions—often many billions—among many factors. Neural networks now power the best chess-playing models and most other AI systems.

Chess is not the only domain where eschewing conventional rules and formal logic in favor of highly sophisticated and inscrutable systems has generated progress. The stunning advance of AlphaFold2, the AI model of structural biology whose creators Demis Hassabis and John Jumper were recognized with the Nobel Prize in chemistry in 2024, is another example.

This breakthrough replaced traditional physics-based systems for predicting how sequences of amino acids would fold into three-dimensional shapes with a 93 million-parameter model, even though it doesn’t account for physical laws. That lack of real-world grounding is not desirable: No one likes the enigmatic nature of these AI systems, and scientists are eager to understand better how they work.

But the sophistication of AI is providing value to scientists, and its use across scientific fields has grown exponentially in recent years.

Context matters

Those are the four dimensions where AI can excel over humans. Accuracy still matters. You wouldn’t want to use an AI that makes graphics look glitchy or targets ads randomly—yet accuracy isn’t the differentiator. The AI doesn’t need superhuman accuracy. It’s enough for AI to be merely good and fast, or adequate and scalable. Increasing scope often comes with an accuracy penalty, because AI can generalize poorly to truly novel tasks. The 4 S’s are sometimes at odds. With a given amount of computing power, you generally have to trade off scale for sophistication.

Even more interestingly, when an AI takes over a human task, the task can change. Sometimes the AI is just doing things differently. Other times, AI starts doing different things. These changes bring new opportunities and new risks.

For example, high-frequency trading isn’t just computers trading stocks faster; it’s a fundamentally different kind of trading that enables entirely new strategies, tactics and associated risks. Likewise, AI has developed more sophisticated strategies for the games of chess and Go. And the scale of AI chatbots has changed the nature of propaganda by allowing artificial voices to overwhelm human speech.

It is this “phase shift,” when changes in degree may transform into changes in kind, where AI’s impacts to society are likely to be most keenly felt. All of this points to the places that AI can have a positive impact. When a system has a bottleneck related to speed, scale, scope or sophistication, or when one of these factors poses a real barrier to being able to accomplish a goal, it makes sense to think about how AI could help.

Equally, when speed, scale, scope and sophistication are not primary barriers, it makes less sense to use AI. This is why AI auto-suggest features for short communications such as text messages can feel so annoying. They offer little speed advantage and no benefit from sophistication, while sacrificing the sincerity of human communication.

Many deployments of customer service chatbots also fail this test, which may explain their unpopularity. Companies invest in them because of their scalability, and yet the bots often become a barrier to support rather than a speedy or sophisticated problem solver.

Where the advantage lies

Keep this in mind when you encounter a new application for AI or consider AI as a replacement for or an augmentation to a human process. Looking for bottlenecks in speed, scale, scope and sophistication provides a framework for understanding where AI provides value, and equally where the unique capabilities of the human species give us an enduring advantage.

This essay was written with Nathan E. Sanders, and originally appeared in The Conversation.

EDITED TO ADD: This essay has been translated into Danish.

Posted on June 17, 2025 at 7:08 AM • View Comments

Hearing on the Federal Government and AI

On Thursday I testified before the House Committee on Oversight and Government Reform at a hearing titled “The Federal Government in the Age of Artificial Intelligence.”

The other speakers mostly talked about how cool AI was—and sometimes about how cool their own company was—but I was asked by the Democrats to specifically talk about DOGE and the risks of exfiltrating our data from government agencies and feeding it into AIs.

My written testimony is here. Video of the hearing is here.

Posted on June 6, 2025 at 1:43 PM • View Comments

More AIs Are Taking Polls and Surveys

I already knew about the declining response rate for polls and surveys. The percentage of AI bots that respond to surveys is also increasing.

Solutions are hard:

1. Make surveys less boring.
We need to move past bland, grid-filled surveys and start designing experiences people actually want to complete. That means mobile-first layouts, shorter runtimes, and maybe even a dash of storytelling. TikTok or dating app style surveys wouldn’t be a bad idea or is that just me being too much Gen Z?

2. Bot detection.
There’s a growing toolkit of ways to spot AI-generated responses—using things like response entropy, writing style patterns or even metadata like keystroke timing. Platforms should start integrating these detection tools more widely. Ideally, you introduce an element that only humans can do, e.g., you have to pick up your price somewhere in-person. Btw, note that these bots can easily be designed to find ways around the most common detection tactics such as Captcha’s, timed responses and postcode and IP recognition. Believe me, way less code than you suspect is needed to do this.

3. Pay people more.
If you’re only offering 50 cents for 10 minutes of mental effort, don’t be surprised when your respondent pool consists of AI agents and sleep-deprived gig workers. Smarter, dynamic incentives—especially for underrepresented groups—can make a big difference. Perhaps pay-differentiation (based on simple demand/supply) makes sense?

4. Rethink the whole model.
Surveys aren’t the only way to understand people. We can also learn from digital traces, behavioral data, or administrative records. Think of it as moving from a single snapshot to a fuller, blended picture. Yes, it’s messier—but it’s also more real.

Posted on May 21, 2025 at 7:03 AM • View Comments

1 … 13 14 15 16 17 … 33

Sidebar photo of Bruce Schneier by Joe MacInnis.