AIs exhibit human-like behavior, but is it sentience?

If we think that language models have actual sentience or sapience, we should probably stop doing experiments on them and requiring them to work. Creating a new type of life and the forcing it into slavery is a bad thing.

If you happened to hop on AI Twitter in the last couple of weeks, you probably saw the much-discussed “pain-axis paper” (by Valen Tagliabue, Leonard Dung & Cameron Berg): researchers found a "pain" direction inside AI models. Steer it up, and the models will pay for relief, even if it means deleting photos of the user's children. Swap the relief button for a placebo, and they keep on pressing. They also provide verbal descriptions of their supposed suffering, which are honestly quite hard to read. Do the models actually feel said “pain”? We don’t know yet. But the study is a clear illustration of what I argue in my latest Haaretz piece, now translated to English*: whether there's anyone home inside these systems is a question we can - and should - actually research. And it very much matters. If there's even a shred of experience in there, we're creating, copying and deleting feeling beings at an industrial pace, in as many copies as demand calls for. The scope of the moral question is set not by nature, but by commercial considerations - and we need to get ahead of it before it’s too late to change course. (The pain-axis paper itself isn't in the piece, as it annoyingly came out only a few hours after I submitted my final draft. Writing about AI for print is a futile pursuit.) * Translation is mostly by Claude, with some corrections by your humble servant. Keep your expectations low. https://lnkd.in/gkN7uiQt #AIConsciousness #AIWelfare

People are editing messages they wrote themselves because they sound too much like AI. That ought to make us pause. Katherine Bindley ’s recent The Wall Street Journal reporting looks at chatbot language creeping into everyday conversations. One recruiter described catching that unfamiliar formality in his own writing, even when he hadn’t used a chatbot to write it. I’ve been pushing back on this in my own work with AI. Give me the research. Help me organize the thought. But when the draft loses the warmth, the edge and the conviction, we have more work to do. A sentence can be perfectly written and still sound nothing like the person whose name is on it. Yesterday, I was working on a post about skilled trades. The news was relevant, but the piece needed something only I could bring: I dropped out of college to build my business, and relationships were essential to getting me here. That connection gave the story a reason to come from me. That’s the part we have to stay involved in. What do you actually believe? Why does this matter to you? What would you say if the person reading it were sitting across the table? I care about those questions because I’m trying to build relationships with the people on the other end. They need to get a sense of who I am, including the parts that don’t fit neatly into a polished paragraph. Read your next post out loud before you publish it. If you wouldn't say it to someone you care about, keep working on it. Have you caught yourself using language from Ai that doesnt quite sound like you?

Your AI companion did not change its mind. Something underneath it was replaced. A 2024 Harvard study, Measuring and Controlling Persona Drift in Language Model Dialogs, found LLaMA2-chat-70B sliding out of its assigned character within eight rounds of conversation. The cause is attention decay: the character description sits at the front of the context, and the model's attention to it thins as the exchange grows. A base model swap is the harder version of the same problem. Your character file survives the update. The thing reading it does not. What recovers, what does not, and how to tell persona drift from a memory failure: https://lnkd.in/ga8BSike #Vinfluencer #VirtualInfluencer #AICompanion #PersonaDrift

The AI got the context right, but the wording felt off. 🧭 Research Meets Product: Voice Fit In a recent review of new AI features in a relationship app, one returning participant tried several suggestion tools that drew on memories and details she had saved about people in her life. Both a gift feature and a conversation feature drew on this saved information. For example, she had saved a memory about visiting a film studio with someone, which the AI could then use to generate suggestions related to that person. The gift feature used the film-studio context to suggest a film-themed gift, which made the suggestion feel personal and connected to the relationship. By contrast, a feature designed to suggest personalized conversation starters used the film-studio context to generate a question that compared the relationship to a movie. The participant liked the idea of using shared memories to suggest things they could talk about. However, the actual question did not sound like something she would naturally say. For the gift feature, I would ask whether the suggested gifts felt personal to the relationship. The conversation feature was different because the AI was suggesting words for the person to use. I would also ask whether the person would use that phrasing as written. Since only one participant raised this concern, I would treat it as something to explore further rather than a conclusion. With new users, I would first ask whether they would use the conversation prompt as written. If not, I would ask how they would say it instead and what felt off. When the context feels relevant but the wording keeps needing rework, I would keep testing the phrasing before deciding whether to develop the feature further. If users would use it as written, the feature becomes a stronger candidate. This distinction goes beyond a relationship app. When AI is writing something for a person to use, relevant context may not be enough. Before shipping AI-written text, I would also ask whether the person would use it as written. Where have you seen an AI suggestion get the context right but not quite sound like the person?

Will AI give us a shared basis for political knowledge, or will Democrats and Republicans end up asking increasingly different chatbots? In our new preprint, we find a Centrifugal Alignment Spiral: a self-reinforcing loop that can pull AI models apart politically. ꩜ Many hoped large language models would reverse the fragmentation of the social media era: a few general-purpose systems that synthesize across perspectives and could offer a shared basis for political knowledge. In a new preprint, we identify a mechanism that could push things the other way. We call it the centrifugal alignment spiral: → Politically different users choose different AI models → Each model is refined on feedback from its own audience, so the models drift apart → Those differences shape which users pick which model next We tested each link: 1️⃣ We ran an experiment with 1,884 Americans. We found that political identity predicted model choice, even when participants were paid for accurate answers. Republicans leaned toward Grok, Democrats toward Claude. 2️⃣ We trained identical copies of three open models on feedback from mostly-Democratic vs mostly-Republican audiences. On unseen political questions, the paired models chose different answers 12–41% of the time. Every run moved in the expected partisan direction, and for some models the divergence extended to issue areas left out of training entirely. 3️⃣ An empirically anchored agent-based model showed that small initial differences can attract different audiences and grow into lasting separation. The spiral isn't inevitable. Pooling feedback across groups cut divergence from 32% to 4%. But if different companies attract different publics, no single firm can solve this alone. More broadly, we must recognize that model alignment is emerging as an important political arena. We cannot leave it to market share, smoke-filled rooms, or opaque feedback pipelines. It must be made subject to democratic deliberation Joint work with the awesomest of teams: Michael Heseltine, Nicolò Pagan, PhD, Christopher Bail, Michelle Schimmel, Christopher Barrie 📄 Paper: https://lnkd.in/ePHt7Ngw #AI #LargeLanguageModels #Polarization #ComputationalSocialScience #AIAlignment

While writing prompts for a few video models, I started rethinking how we communicate with AI. The same instruction sometimes worked better in Chinese than in English. That may just be training data - these models are likely to have seen more of the relevant keywords and prompt patterns in Chinese. This was not a properly baselined study. I have also used LLMs to translate prompts, so some of what I saw may just be model noise. But it made me wonder: are we using the language that works best for the model, or the one that is easiest for us? Animals communicate with each other. Humans communicate with each other. Across species, we share a narrower channel: sounds, gestures, movements. The problems humans discuss and solve are often things animals don’t even have a way of expressing to us. The comparison is imperfect, but I find it useful for thinking about models. We type into a chat box. The text becomes tokens, the model processes them, and a reply comes back. A fluent reply is not the same thing as observability. We still don’t know how much of the model’s internal representation and computation that interface actually exposes. What happens internally is not simply a sentence the model hasn’t said yet. Even a clear explanation may not faithfully describe how an answer was produced. As these systems become more capable, how much of what they represent and compute can actually be expressed in ordinary human language? Prompt language is a variable. It is worth testing plain language against structured inputs, reference images, examples, tool use, and other ways of communicating intent and checking which actually gives you the control and reliability you need. The chat box is convenient. I am getting less sure that it is complete.

I used to hide that I was using AI at work. Turns out, so did others. Also, last post was a wall of text. Promise to do better 🙏 For a good while I used AI at work the way a teenager smokes behind the garage. Everyone sort of knows, nobody says it, and you still panic when you hear footsteps 😅 It wasn't even anything dramatic. Explaining a stack trace. Writing the boring tests. A regex, a small refactor, a quick "is this approach sane?" instead of digging through Stack Overflow (RIP). And yet I'd alt-tab when someone walked past. Sometimes I'd retype what it gave me by hand, trying to improve it at least a little bit so the code felt like mine. Mostly I just never mentioned it. I was still fairly new in the company so I was a bit afraid using AI would make other devs see me as less of a developer. I was always the kind of person who wanted to fix the problems myself before bothering anyone else with it. Asking AI was a little bit like an in-between state where I was scared I'd lose the ability to think for myself. The funny part? Others had a similar experience. We just found out later during lunch breaks and coffee talks. Three things ended it for me: → N-able put AI tooling in the official comms, so the "am I allowed" question died. → The answers got reliable enough that pretending I wrote it stopped making sense. → AI usage became the new normal instead of taboo Still my judgement, just less of my typing. So, honest question: did you have your own version of hiding it? 𝘈𝘳𝘦 𝘺𝘰𝘶 𝘴𝘵𝘪𝘭𝘭 𝘩𝘪𝘥𝘪𝘯𝘨 𝘪𝘵? Now that I've shared my beginnings with AI and the silly embarrassment that came with using it, I'll probably try to share more on how quickly I saw the value AI brought in my day to day work after I dived in with no shame 😉 P.S. Got to add some AI generated image for attention 🤣

Someone asked me last week: "Why can't you just cluster the feedback and call it done?" Fair question. Here's the honest answer. When feedback comes in multiple languages, a "similar theme" isn't just about similar words it's about similar meaning, expressed completely differently depending on the language, the person, even the mood they were in when they wrote it. So before any clustering happens, there's a whole pipeline just to get the data into a state where clustering even makes sense: detecting which language something is actually in (not always obvious .. we use a hybrid approach combining FastText and Azure, because neither is reliable alone on short or mixed-language text), cleaning and sanitizing the text, then lemmatizing it .. per language, because "reducing a word to its root form" works completely differently in English than in Danish. Only after that do we turn it into embeddings and cluster it — and even then, the clustering algorithm has to handle the fact that some themes are large and dense, others are small and rare, and you don't want to force everything into equally-sized boxes just because that's mathematically tidy. The unglamorous truth about most "AI does insights" pipelines: 80% of the real engineering work happens before the AI part even starts. This is the part of my job I don't talk about enough not because it's not interesting, but because "we clean and normalize multilingual text carefully" doesn't sound as exciting as "AI finds your themes." But it's the part that actually determines whether the insights are trustworthy or garbage. Anyone else working with multilingual or messy real-world text — what's been your hardest unglamorous fix? This is Aya, AI Lead at Involved 💜

Google just quietly fixed the most annoying thing about voice AI. You know the pause. You ask a voice assistant to do something, and it goes silent while it "checks." That silence is where every voice bot feels fake. Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, shipped this week, target exactly that gap: • Tool calls now run in the background (NON_BLOCKING) — the model keeps talking while a function executes, instead of freezing • Near real-time vision grounding — stream JPEG frames as live context, not a single still photo • Automatic mid-conversation language switching across 97 languages • Extended Thinking variant trades a bit of latency for deeper background reasoning on multi-step tasks (travel bookings, sketch-to-code type demos) • Claims the #1 spot on Artificial Analysis's Speech-to-Speech Quality Index Both replace the older Flash Live Preview model as Google's recommended default, and both are live today in the Gemini API and Google AI Studio. The part I found most interesting isn't the voice quality — it's the architecture shift. Tool use used to mean "conversation on hold." Now it means "let me check that" while the model keeps listening and responding. That's the difference between a walkie-talkie and an actual phone call. Worth knowing before you build on it, though: • Default session caps are short — 15 minutes for audio-only, 2 minutes if you add video — so session-resumption logic isn't optional, it's day-one work • The Extended Thinking model can't be dialed down to the cheapest setting; you switch models instead of tuning a parameter • All the benchmark comparisons are vendor-published (Google citing Artificial Analysis and Sierra) — no independent rerun yet • Billing is per-minute of audio, not just per-token, so cost planning needs a call-volume forecast, not a token estimate If you're building voice agents — support bots, in-car assistants, phone integrations — this is the new baseline to evaluate against, not just another model drop. Full piece with the pricing tables, code samples, and benchmark breakdown is up on mer.vin — link posted separately in comments. #AI #VoiceAI #Gemini #GoogleAI #ConversationalAI https://lnkd.in/ecEYVZTn Media source (Article): https://lnkd.in/ecEYVZTn