Marc-Alexandre Cartiant (@macartiant) header
Marc-Alexandre Cartiant (@macartiant) profile picture

At the sweet spot between Management, Digital Marketing and Engineering. Bullish about LLMs from 2022.

Japan
Joined October 2013
Photo by Marc-Alexandre Cartiant (@macartiant)Photo by Marc-Alexandre Cartiant (@macartiant)Photo by Marc-Alexandre Cartiant (@macartiant)Photo by Marc-Alexandre Cartiant (@macartiant)Photo by Marc-Alexandre Cartiant (@macartiant)Photo by Marc-Alexandre Cartiant (@macartiant)Photo by Marc-Alexandre Cartiant (@macartiant)Photo by Marc-Alexandre Cartiant (@macartiant)Photo by Marc-Alexandre Cartiant (@macartiant)Photo by Marc-Alexandre Cartiant (@macartiant)
Asking long-time users of complex internal systems (ERPs, core banking, claims platforms, etc.) to test AI tools and report productivity gains after one or two workshops produces noise, not signal. These users know the current process deeply. They are rarely positioned to redesign it. 1. Value creation is routinely put to the wrong audience ERPs and automation were not designed by asking every end user “how would automation make you more productive?” Users had a critical role in surfacing pain, exceptions, and failure modes. They did not own the system architecture, the process redesign, or the decision on how much control to retain. Treating them as the primary architects of transformation produces local optimisations and missed systemic gains. 2. The right conversation involves three groups with distinct jobs: - Users diagnose friction, volume, exceptions, and quality issues. They are excellent sensors. - Process and risk owners decide what must stay human, what can be assisted, and what can be automated under which controls. - Leaders with P&L or operating accountability choose the degree of change the organisation will actually absorb and fund. Asking only the first group produces anecdotes. Asking only the third produces paper strategies. The value sits in the structured interaction between them. 3. Do not confuse training with transformation A minimum level of hands-on capability is necessary. Just as organisations train people on Excel, they will need to train them on AI tools. The distribution of power users versus casual users will look similar to traditional office software. Creating the ability to use a tool and redesigning the work are separate workstreams with different owners, timelines, and success criteria. Conflating them is how companies declare victory after a workshop series while the underlying processes remain unchanged. Bottom line Execute the brilliant basics by articulating the right problem statement. Only after that should an organisation decide which tools to buy or build, which steps to automate, which decisions stay human, and what training is required to make the new design work. The inverse — starting with tools and looking for problems — produces activity without impact.
2
2
114
The screen displays an error message. I try a command; things get worse. More messages, just as alarming, scroll by rapidly. I start to panic. Whatever I do, I feel like I’m breaking something. I lose track of the time that passes between each attempt and each failure. Then my vision goes completely dark… even though my eyes are open. I’m in my room. The first question that comes to mind is: “Am I offline?” Then: “Did all of this happen only in my head, without changing anything in reality?” I realize I’m fully back in the real world, and I think to myself that Nature built us in such a way that what happens inside our heads — especially when we dream — leaves us “offline,” completely disconnected from the world. But for how much longer? Cybernetic implants grafted onto the brain to compensate for a disability will one day be standardized. The same implants, used this time to gain productivity, are probably already at the stage of concrete experiments — or worse: wars unfortunately serve as pretexts to develop and deploy these technologies, even as prototypes, directly in the field. The question then arises: will these implants need to remain functional while we sleep? If so, we will no longer truly be “offline.” Whatever comes out of our heads — even in the form of a dream — will have an impact, probably negative, on our environment. Nature imposes needs on us, biological necessities for survival. It also offers ways to circumvent them: pills to delay sleep, or simply caffeine. If we already know how to leverage Nature to create objects that satisfy or divert these necessities, should we imagine that one day sleep itself will become another moment in our daily cycle meant for producing, for creating value — in other words, for working?
2
35
A generalist AI model looks like a service. Architecturally, it is something much stranger: a broadly capable component whose reach must be deliberately constrained. Enterprise systems are built from parts with contracts: defined inputs, defined outputs, replaceable internals, and a way to fail without taking the rest of the system with them. A generalist model breaks that assumption. The same model can summarize, code, interpret a contract, generate financial analysis, or invent an answer that convincingly resembles any of them. Capability is the point. Unbounded behavior is the integration problem. So the model can stay general. The component around it cannot. 1. Capability is not authority Prompting expresses intent. Architecture enforces it. Decide, in software, which data the model can see, which actions it can take, and which outputs are invalid until a deterministic check says otherwise. If the boundary lives only in the prompt, you do not have a component. You have a hope. 2. If you cannot pin it, test it, and roll it back, you do not own the process Enterprises do not run on “the model got better last Tuesday.” They run on versions, support windows, regression suites, migration windows, and rollback mechanisms that restore prior behavior — including across layers a provider may change without renaming the model. Change has to happen on your lifecycle, not the provider’s. 3. If you cannot measure failure on this task, you are trusting a demo Not a benchmark. This workflow. These conditions. This cost of being wrong. How is quality measured? How is bad output detected? What happens when performance drops below the line? Small teams can buy much of that wrapper from a vendor, and often should. Critical, regulated, or high-consequence systems cannot outsource responsibility for the wrapper. The wrapper is the responsibility. The model may remain enormous, flexible, and generalist. The system around it cannot.
🤖 Made with AI
24
Most companies are staffing AI-augmented support the wrong way. They’re either dumping it on traditional helpdesks that have barely been reskilled, or burning expensive AI specialists on tickets that don’t need them. One creates permanent friction. The other burns out the people you actually need for harder problems. Every major productivity platform now ships with AI that can either do the work or walk the user through it. Routine how-to questions are increasingly Tier 0. The AI handles them, or it should. The real friction sits one level up. Users still need human help. Most organisations treat this like a one-off training project instead of a continuous support model: a few hours of modules, a prompt list, then silence. Followed by creative metrics to prove the investment worked. What’s required is a three-layer structure: Tier 0 - AI Handles the majority of basic questions and a growing share of the actual work. Tier 1 - Human support for the new way of working The critical layer. Users come here when the AI fails, when they don’t know how to approach the task, or when they need help shifting how work gets done. This should evolve from existing desktop and productivity support teams, not ultra-specialised AI experts. Tier 2 - AI & business specialists Only for workflow redesign, automation, deeper domain expertise, or advanced AI capabilities. Two realities make this harder than it looks. The boundary between tiers keeps moving. What needed a specialist last quarter is routine this quarter. Static models lag. And the usual approach to “upskilling” is theatre. Real capability comes from continuous practice on live work, feedback from actual tickets, updated expectations, and the mandate to help users change how they work. Some existing staff will thrive. Some will not. Governance, data quality, and security sit underneath every tier. Weak foundations amplify the friction. Give existing support teams the skills, tools, and mandate to own the middle layer. Keep specialists on problems that actually require them. Because if you hire highly specialised AI talent and relegate them to low-value support… don’t be surprised when they leave. The tools have moved. The support model has to move with them - continuously. Authored with @DavidHardoon
1
71
Agentic AI Risks in 60 seconds. Who's in control? Watch 👀 Co-authored with @DavidHardoon
53
The real value in complex corporate banking deals almost never lives neatly in the CRM. It’s scattered across email threads, marked-up PDFs, chat messages, covenant negotiations, and those unofficial Excel trackers everyone quietly relies on. Sound familiar? One email has the client’s latest pushback. A PDF holds the risk team’s comments. Someone’s personal spreadsheet shows the actual status. Meanwhile, the CRM is politely outdated. The result? Reviews drag on, decisions get delayed, and teams burn hours on inbox archaeology. Most banks are wasting millions on AI pilots while still drowning in email archaeology. Dumping every scattered file into a chatbot usually creates more confusion than clarity, especially inside regulated banks. But there is a practical, compliant approach that works: Step 1: Create a controlled “Golden Source” Use AI to help build and maintain one governed reference document, always with human approval before anything becomes official. This could be a clean Excel tracker, deal summary, or living document in secure SharePoint/Microsoft 365. At minimum, capture: - Facility / Client reference - Key covenants & terms - Client feedback summary - Open actions, owners & deadlines - Risk flags - Last updated + source references Step 2: Make the Golden Source the AI’s single source of truth Instruct every AI tool to answer only from this document and cite its sources. No hallucinations. From there you can instantly generate draft call memos, credit summaries, action lists, and clear “What’s the real state of this deal?” overviews. Human review stays mandatory. Version control stays mandatory. CRM still gets updated. The difference? Your team finally works from one trusted operational view instead of memory and spreadsheet folklore. Banks don’t just need better AI. They need a governed foundation first. The hard part isn’t building the Golden Source. It’s making one that actually sticks instead of becoming another abandoned SharePoint graveyard. Co-authored with @DavidHardoon
1
75
Every week, I receive LinkedIn messages pitching the latest “agentic” AI products. And almost every time, I have the exact same reaction: This is going to be brutally hard to sell into a large enterprise. Not because the ideas are bad — but because the market has fundamentally changed. Microsoft Copilot + Power Platform, Claude for Work, and Google Workspace + Gemini have raised the bar so high that many “agentic” tools now feel like expensive add-on features rather than standalone solutions. You’re no longer just competing with other startups. You’re competing against: • Their massive distribution • Deep native integrations • Enterprise-grade security and compliance • Internal champions already embedded inside the account That’s a tough starting point. Before pitching any large enterprise, honestly ask yourself these 4 non-negotiable questions: 1. Is your product truly differentiated? Not “better UX.” Not “more flexible.” Not “AI-native.” Can it actually do something the major platforms cannot do well enough inside a real enterprise stack? If the answer is no, most conversations die before they even start. 2. Do you have real paying customers? Not pilots. Not design partners. Not vanity logos. Actual companies paying real money and getting measurable value. Enterprise buying is never just about the license fee — it’s about security reviews, compliance, legal, procurement, change management, and internal politics. Those hidden costs often dwarf the sticker price. 3. Can you support messy, real-world integration? Enterprise onboarding is rarely clean or quick. You don’t need standard customer success. You need senior people who can live inside the customer’s architecture, processes, politics, and edge cases for weeks or months. Without that capability, your tool often gets quietly decommissioned after 6–12 months. 4. Why should they buy instead of build? This is the killer question. At some point, the CIO, CDO, or Head of Automation will ask: “Why can’t we just build this ourselves?” Your answer must be economically compelling, technically credible, and risk-aware. Vague hype won’t survive that conversation. My advice to anyone building in the agentic AI space: 1. Stop chasing large enterprises too early. Unless they’re ready to become strategic partners or investors, protect your runway. 2. Start with ambitious mid-market companies and growth-stage SMEs. This is where you reach real decision-makers faster, close actual deals, get raw feedback, build strong references, and improve your product in the real world. 3. Stop treating startup conferences and tech events as primary sales channels. Great for visibility and networking — usually terrible for finding customers who will actually pay. The agentic AI market is maturing fast. The pure hype phase is over. ... And for those who are trying to do the above while simultaneously pursuing market penetration in Japan... then pass — you're not ready yet. --- post co-authored with @DavidHardoon
1
100
The Oversight Bottleneck: Why AI Code Generation Raises the Bar of Human Responsibility We now have multiple ways to code. We can still write every line ourselves. We can write the structure and let the model fill in the details. Or we can describe the intent at a high level and let successive generations of code accumulate into a working system. But in every case, the human remains the initiator — and the final arbiter. The same pattern is appearing in other domains that produce complex, high-stakes artifacts: - contract drafting, - architectural and engineering design, - financial structuring. The mechanism looks similar: - A human provides intent or partial work. - The system generates volume and complexity. But the verification regimes are very different. - A contract is tested in negotiation or court. - A building is tested by gravity. - Code is tested by compilers, test suites, production traffic, security review, and users. So the generalization is only partial. AI changes the economics of production across these fields. It does not remove human accountability. The visible limit appears when the generated artifact exceeds the human’s ability to review it meaningfully. At that point, the old model of “generate, then review” starts to break. The model can iterate on its own output. It can fix inconsistencies. It can refactor without being reminded of the original constraints. That autonomy is useful — until it isn’t. Because the person who must sign off, deploy, approve, or take legal responsibility may no longer be able to verify that the output implements the intended specification. Responsibility does not disappear. It becomes harder to exercise. This is not solved by telling developers to “just read the code.” When the codebase is large, generated quickly, and architecturally unfamiliar, full human comprehension is no longer a realistic control mechanism. The bottleneck is no longer generation speed. It is review and judgment capacity. And this is the important point: the skills required for effective oversight are not fundamentally different from those required for high-quality original work. To judge whether generated code, a contract, or a financial structure implements intent, respects constraints, avoids hidden failure modes, and remains maintainable, you still need deep expertise. AI does not deskill the practitioner. It changes the unit at which expertise is applied. The people who treat AI as magic that also magically solves oversight will eventually rediscover an old rule: You own what you ship.
95
I have never seen a no-code or low-code platform that can truly replace real code for genuinely complex systems. We are closer than we were ten years ago, but one stubborn reality remains: the cost of understanding, evolving, and governing a large software artifact at scale still requires serious engineering skill and discipline. The fantasy of fully specifying software without code is ancient. The Gang of Four era turned UML into near-religious practice: model first, code later. In theory it promised clarity and reuse. In practice, outside of database schemas (where the model is executable), those diagrams almost always became either superficial illustrations or expensive liabilities that diverged from reality the moment implementation began. High-level diagrams can usefully sketch primary flows and occasionally follow standards. They are, however, almost always incomplete snapshots of intent rather than authoritative specifications. They are cheap to produce when someone is motivated and expensive to maintain when reality shifts. Reverse-engineering an existing codebase into clean models, or starting from models and keeping them in sync, quickly reveals the friction. In environments that demand deep governance, that friction is only justified if the payoff in auditability or onboarding is massive. In most cases, the code itself remains the single source of truth precisely because it is executable, testable, and versioned. We have already successfully raised the abstraction level many times. We no longer write processor instructions by hand. We trust compilers, runtimes, frameworks, and libraries. The question is not whether further abstraction is possible, but whether the next layer (visual builders, executable models, or AI-generated systems) can deliver the same combination of power, observability, and control that code currently provides. The decisive variable is the required level of governance. In safety-critical, heavily regulated, or long-lived systems, you may need explicit traceability, formal properties, or generated artifacts with full audit trails. In those contexts, investing in stronger modeling or constrained languages can make sense. For the majority of commercial software, however, the pragmatic answer remains the same: well-structured code, comprehensive automated tests, clear architectural decision records, ruthless review processes, and continuous integration deliver sufficient governance at lower long-term cost than heavyweight modeling approaches. The future will not be “no code.” It will (actually it is already) be less code written by hand, with engineers operating at a higher level through AI copilot and tools that keep intent and implementation in sync automatically. But code remains both the means of production and the ultimate artifact we must govern.
24
Fix the Input. Then Amplify. Most organizations don't fail because they lack ambition, talent, or technology. They fail because they amplify broken inputs: - A vague strategy. - A misaligned incentive system. - Poor data. - Unclear decision-making. - Conflicting objectives. Then they add more people, more budget, more process, or more AI. The result? They don't solve the problem. They scale it. 1. The mistake Too many initiatives start with goals like: - "Be more innovative." - "Improve customer experience." - "Leverage AI." These sound ambitious, but they're not operational. When objectives are fuzzy, systems optimize for activity instead of outcomes. You get: ✅ More meetings ✅ More dashboards ✅ More automation ❌ Better decisions 2. The highest-leverage move Before scaling anything, fix the input. That means: - Define the problem in observable terms. - Clarify what success actually looks like. - Align incentives and decision rights. - Build feedback loops that are difficult to game. - Eliminate objectives you cannot execute well. This isn't strategy. It's organizational hygiene. And it's often the most neglected work because it isn't glamorous. 3. Where AI fits AI is neither a savior nor an existential threat. It's an amplifier. When goals and constraints are clear, AI can dramatically increase speed, scale, and execution quality. When goals are vague, AI simply produces larger volumes of confusion at greater speed. The best human-AI systems follow a simple pattern: - Humans define and refine the problem. - AI scales the execution. Conclusion: Fix the input. Then amplify. Not because you want to move slowly. Because it's the fastest way to move in the right direction. The real choice isn't between ambition and caution. It's between: - Amplifying quality - Amplifying dysfunction And one of those compounds much better than the other.
15
Can I trust this AI not to deliberately mislead me, slow me down, or steer my reasoning in a hidden direction? Today, when an AI produces a problematic answer, it is difficult to know what we are looking at: - Is it a technical error? - A hallucination? - A safety mechanism? - A political constraint? - A commercial filter? Or a deliberate design choice shaping what the model is allowed to say? These things are often mixed together, but they are not the same. A hallucination is a reliability problem. A hidden steering mechanism is a governance problem. And in some cases, it becomes a sovereignty problem. We already see this clearly with some state-controlled models. Certain answers are restricted not because the model is technically unable to answer, but because it has been designed not to cross political boundaries. That example is obvious. The harder question is what happens when similar forms of steering are less visible, presented simply as “model behavior,” “alignment,” or “safety.” AI reasoning is never fully neutral. It reflects human design choices: technical, ethical, commercial, legal, and sometimes geopolitical. So the real question is not whether AI systems are influenced. They are. The question is: by whom, for what purpose, and with what level of transparency? This is why I believe the next wave of AI governance will not only focus on accuracy or benchmark performance. It will also focus on: - reasoning integrity - auditability - independence from external influence - transparency of constraints - resilience against political or commercial capture - technological sovereignty In the future, organizations may not only ask: “Which model is the most powerful?” They may ask: - “Which model can we trust?” - “Which model can we audit?” - “Which model keeps our reasoning process under our control?” The ability to prove that an AI system is reliable, transparent, and free from hidden external constraints may become a major competitive advantage. More importantly, it may become a strategic necessity. Because the real risk is not only that AI gives us wrong answers. The deeper risk is that we no longer know who is shaping the answers we receive.
29
AI in HR Tools: From Form-Filling to Systems That Actually Help Careers Most current “AI-powered” HR tools for career management and performance evaluations I have seen are optimizing the wrong thing. Many of them are very good at one task: helping employees turn self-reported objectives, achievements, and development areas into polished text, suggested metrics, or ambitious-sounding goals. That can feel productive. But if the output still feeds into a process that managers skim, HR archives, and few people reference when making promotion, compensation, or role decisions, then we have not meaningfully improved the system. We have simply made low-value administrative work faster. The more important question is this: What would it take for AI-augmented HR systems to create real value for both employees and organizations? I think three shifts matter most. 1. From self-reporting to evidence synthesis Instead of asking employees to recreate their year from memory, AI should help synthesize verifiable signals: Project outcomes. Peer feedback. Stakeholder input. Delivery metrics. Demonstrated skills. The employee’s role should be to validate, contextualize, and correct the synthesis — not perform creative writing. 2. From one-sided forms to shared visibility One of the biggest missed opportunities is giving employees and managers access to a common, appropriately permissioned view of performance signals. Not surveillance. Not raw activity tracking. Not black-box scoring. But synthesized insights such as: - “This person consistently creates outsized impact in cross-functional projects.” - “Multiple stakeholders mention strong execution under ambiguity.” - “Feedback suggests a recurring development opportunity around delegation.” When both sides can see the same evidence base, performance conversations become less subjective and more diagnostic. 3. From generic goals to development contracts The highest-value use case is not generating better wording for objectives. It is turning past performance data into forward-looking development recommendations. What skills does this person need for their next step? Which roles fit their demonstrated strengths? What experiences would help close the gap? What should the manager and organization commit to providing? That changes the performance review from a backward-looking rating exercise into a forward-looking development contract. And that is where AI can become genuinely useful. The goal should not be to make HR forms faster to complete. The goal should be to improve the quality, transparency, and actionability of the information that actually shapes career decisions — while keeping humans responsible for judgment, context, and accountability. Today, many AI HR tools optimize for the appearance of innovation. The next generation should optimize for clarity, trust, and better career conversations. That is the version worth building.
37
I now possess sufficient empirical evidence to declare an uncomfortable truth: computers can smell fear. Try to print a critical report at 11:47 p.m. the night before a deadline. The printer, previously silent and cooperative, suddenly begins speaking in its own ancient dialect of blinking lights and cryptic error codes. Translation: “Dream on, human.” Paper jams materialize in compartments that have no business jamming. Ink levels drop from 80% to “replace cartridge immediately” in the span of three pages. The email server is in on the conspiracy. Messages labeled URGENT in all caps, decorated with red exclamation marks and threatening emojis, sit patiently in the inbox all day—until the exact moment you sit down to answer them. At that precise second, the server sighs, shrugs, and delivers a polite but firm error message. The more important the email, the higher the probability of digital mutiny. And then there is the AI. During the big demo, in front of stakeholders and polite smiles, the system performs beautifully at first. It displays “Thinking…” with an elegant progress animation. It even offers scientifically flavored reassurances about processing your request with quantum-inspired precision. The audience leans in. You allow yourself a flicker of hope. Then it delivers the coup de grâce: “An error happened. Try again?” Pure, refined psychological warfare. This is Murphy’s Law 2.0 in full effect: "Any system that can fail will fail at the exact moment it will cause you the maximum possible embarrassment, delay, or financial damage." We didn’t just add complexity to technology. We gave it TIMING! Machines have been granted not only processing power but also the exquisite ability to mock their creators at the worst possible moment. They can now smell desperation, deadlines, and that tiny bead of sweat forming on your forehead. That’s when they strike—with elegance, perfect comedic timing, and zero mercy. We were losing this war from the beginning. Before the digital age, at least the betrayals were slower and more honest. A typewriter ribbon would snap. A fax machine would eat the page. Today the betrayal is instantaneous, global, and backed by cloud infrastructure. Yet here we are—still clicking “Retry,” still updating drivers at midnight, still pretending the machines are on our side. They’re not. They never were. Stay vigilant. Lower your expectations. And above all: never let the machines know you have a deadline.
1
34
This is deep
Replying to @MyleneFlicka
I take that same approach to programming. I want to do it, not have it done.
18
Marc-Alexandre Cartiant retweeted
emergent mind
They are ants solving a geometric problem and it is mind-blowingly colorful.
33
66
4
665
61,540
Even Grok is having fun with my prompts...
18
I took a photo from almost 20 years ago and asked Grok to progressively age me an additional 40 years. The result is both amusing and terrifying. Amusing because Grok didn’t really capture how I look today (the ultimate accuracy test). And yet… by the end, I wouldn’t be surprised if it’s actually close. What troubled me most is the very last frame… you’ll understand why if you watch the video. Curious: which age looks the most realistic to you? 👀
35
Screens optimize for speed. Paper optimizes for truth. AI optimizes for leverage. I’ve gone back to writing by hand—real paper, real pen. I’ve almost stopped buying e-books. The difference is brutal. A screen is frictionless. You can edit forever, rearrange instantly, bail the second a sentence fights back. Paper does the opposite. There’s weight. There’s cost. Each word is permanent. No backspace. No undo. Handwriting is cognitive resistance training. Here’s the sequence almost nobody uses: Paper → AI → Network. Paper forces depth. AI pressure-tests it. The network distributes it. Once the idea survives ink (and in French, that is, my mother-tongue), I snap a photo. Seconds later AI transcribes, translates, tightens, attacks every weak joint, and hands it back sharper. Not to replace the thought—to forge it. Use AI first and it dilutes. Use it second and it multiplies. That order is everything. What separates us from machines isn’t the tool. They’ll simulate handwriting by tomorrow if they haven’t already. It’s that we deliberately choose constraint before leverage. Paper for incarnation. AI for refinement. Digital for scale.
24
AI isn’t a babysitter. With kids, it stops being "just a tool". That’s why I put ChatGPT (and similar tools) in the same box as social networks and the open web: my kids don’t get it on their phones. Not because I’m clutching pearls. I’m the same parent as everyone else—checking geolocation more often than I should, enjoying the comfort of “knowing where they are.” So no, this isn’t some anti-tech posture. It’s just me noticing one simple thing: unlimited access + intimacy is a weird mix, especially when the “listener” is a product. We do let them use Instagram or AI tools sometimes, but only on my phone with the family account. Side effect: the platforms start thinking I’m a teenage girl. My Instagram feed becomes K-pop edits and study hacks. And the AI tone shifts too, like it’s talking to someone younger and “studious.” It’s funny… but it also makes the point: these tools don’t just reply. They *build a profile*. Quietly. Another thing I’ve seen in real use: the default setting is often ego-friendly. Give it a half-finished draft and it’s weirdly generous. I’ve tried this across ChatGPT, Grok, Gemini. Unless you explicitly ask it to be harsh, it tends to start with “This is great!” and an 8/10 vibe. Adults can roll their eyes and move on. Kids? That’s different. They’re still forming taste, judgment, and confidence. Flattery lands harder when your sense of self is still under construction. And then there’s the bonding part—because yes, it happens. I’ve been there: late-night deadline, stuck on a prototype, going back-and-forth with the AI until the last blocker falls. You end up thanking it like it’s a colleague. As an adult, I can label that as transactional. But I also see how easy it is for a kid to experience it as something else: a “someone” that’s always available, always calm, always on their side. That’s the part that bothers me most. Not the “answers.” The relationship shape. If a child starts using it as a private place to vent, to ask social questions, to get reassurance, it becomes a kind of shadow counselor. Except it’s not accountable, it doesn’t know your kid, it doesn’t see context, and it doesn’t carry consequences. It just keeps talking. And yes, we’ve already seen lawsuits and public cases where families claim chatbot interactions played a role in serious harm. I’m not pretending every headline is proven fact. But the direction is obvious: when supervision disappears, the edge cases show up fast. So in our house the rule is simple: no AI tools on their phones. If they want to use it, they can do it with me nearby, on my device, like any other part of the wild web. And the “teenage girl profile” thing? I don’t take it personally. It’s just a reminder that these systems mirror whoever is holding the screen—sometimes a little too well.
40