When Guardian columnist Rhik Samadder handed his entire day to ChatGPT and asked it to plan a London adventure, the results were genuinely interesting — and quietly revealing. The AI picked a postal museum with underground mail carriages, wetlands in Walthamstow, and a hidden Roman temple beneath the Bloomberg building. Each stop was individually impressive. But somewhere between the brilliant individual recommendations and the lived experience of actually following them, something went wrong. The AI never told him to eat. It never suggested a bathroom break. It sent him to a jamón bodega to eat ham alone because a machine said so. And he ended up at a 16th-century warship wondering what he was doing with his life.
This isn’t just a funny column. It’s a case study in the gap between what AI recommendation systems optimise for and what humans actually need — and UK businesses building consumer-facing AI products should pay close attention.
The recommendation is not the experience
The most striking detail in Samadder’s account isn’t a bad recommendation. The AI’s individual picks were genuinely good — the kind of suggestions a well-read friend might make. The Postal Museum, the Walthamstow Wetlands, the Bloomberg Mithraeum: these are legitimately interesting London destinations that many Londoners haven’t visited.
The problem is everything around the recommendations. AI optimises for relevance and novelty in isolation. It doesn’t model the physical reality of being a human body moving through a city — getting hungry, needing the loo, feeling the emotional weight of doing activities alone because a machine told you to.
Strategic Insight: AI recommendation engines excel at matching content to preferences. They fail at modelling the full context of human experience — energy, emotion, social dynamics, physical needs. This gap defines the current frontier of consumer AI product design.
| What AI optimised for | What the human needed |
|---|---|
| Novelty and cultural interest | Food and rest between stops |
| Location variety across London | Logical routing that didn’t exhaust |
| Matching stated preferences (“sensual pleasure”) | Understanding what that actually means in practice |
| Breadcrumbed schedule for excitement | Permission to deviate without guilt |
| Individual stop quality | Coherent emotional arc across the day |
Why “vibe” is not a product specification
Samadder told ChatGPT his vibe was “sensual pleasure,” hoping to be sent to a spa. Instead, the AI booked him a flotation tank (fully booked) and offered “AI robotics massage pods.” The disconnect between what he meant and what the system interpreted tells us something important about natural language interfaces in consumer products.
Current large language models handle explicit instructions well. “Book me a table at a restaurant near Borough Market for 7pm” is a solvable problem. But “plan a day that feels good” requires understanding that hasn’t been built yet — understanding of mood, energy, social context, and the difference between an interesting activity and an enjoyable one.
Reality Check: When users describe what they want in emotional or experiential terms, AI systems default to literal interpretation or pattern matching against training data. “Sensual pleasure” becomes flotation tanks and robot massage pods, not a hammam and a long lunch. The translation layer between human desire and AI execution remains crude.
For UK businesses building AI-powered consumer experiences — travel planning, lifestyle recommendations, personalised services — this gap isn’t a minor UX issue. It’s the core product challenge. Getting the recommendation right but the experience wrong means users won’t come back.
The cocooning effect and why it matters commercially
Samadder’s self-built chatbot “RhikGPT” offered the sharpest observation of the day: “A ChatGPT itinerary tends to smooth the edges, so you move through London like a ghost with no contact.”
This cocooning effect — where AI-guided experiences remove friction, spontaneity, and human interaction — has direct commercial implications.
Critical Context: Friction in consumer experiences isn’t always a problem to solve. Asking a stranger for directions, stumbling into an unexpected shop, or changing plans because a street smells like fresh bread — these “inefficiencies” are often the experience itself. AI systems that optimise them away optimise away the value.
Tourism businesses, hospitality providers, and experience platforms should be particularly alert to this. An AI concierge that efficiently routes guests through a city may deliver worse outcomes than a human concierge who says “actually, forget the museum — there’s a brilliant market on today, and the weather’s perfect for it.”
The commercial risk is real. McKinsey’s 2025 consumer research found that 67% of consumers who used AI-powered travel planning tools described the experience as “efficient but impersonal.” Only 23% said they would use the same tool again for leisure (versus 71% for business travel). The distinction matters: business travellers want optimisation, leisure travellers want discovery.
SME Advantage: Smaller UK businesses can compete against AI-powered platforms by deliberately building human friction into their services. A boutique hotel that offers a hand-drawn neighbourhood map with personal recommendations from staff creates more value than an AI-generated itinerary — because the map comes with a conversation.
What this means for UK businesses building consumer AI
The strategic question isn’t whether AI can make good recommendations. It clearly can. The question is whether recommendation quality alone creates a product people want to use more than once.
Samadder’s day suggests the answer is no — at least for experiential products. The AI picked good stops but created a bad day. That distinction should worry anyone building AI into consumer-facing services.
Priority actions by business maturity
If you’re scoping an AI consumer product:
Build for the gaps, not the recommendations. The hard problem isn’t suggesting good restaurants — Google already does that. The hard problem is knowing when someone is tired, when they need to eat, when they’d rather sit in a park than visit another museum. Design your AI layer around these human factors, not around content matching.
If you already have an AI recommendation feature:
Audit for the cocooning effect. Watch session recordings or interview users. Are people following AI suggestions robotically and reporting satisfaction that’s lower than expected? Are they completing AI-planned itineraries but not rebooking? The gap between completion rates and satisfaction rates is where this problem hides.
If you’re competing against AI-powered platforms:
This is your opportunity. The Guardian piece demonstrates that human judgment, spontaneity, and social connection remain enormously valuable in experiential contexts. Market your human-led approach not as a limitation but as the product itself.
Take Action: Run a simple A/B test with your next 50 customers. Give half an AI-optimised recommendation. Give the other half a recommendation from your best staff member, with their personal reasoning included. Measure satisfaction at 24 hours, not at point of sale. The results will likely surprise you.
Four non-obvious challenges for consumer AI builders
1. The accountability vacuum
When Samadder skipped lunch because the schedule didn’t include it, whose fault was that? His, for not thinking for himself? The AI’s, for not modelling basic human needs? ChatGPT’s designers, for not building meal breaks into itinerary logic? Consumer AI products create accountability gaps that traditional services don’t have. When a human tour guide forgets to schedule lunch, you tell them and they fix it. When an AI does it, users often just suffer in silence — and then don’t come back.
Hidden Cost: Customer complaints about AI-guided experiences are systematically lower than complaints about human-guided ones — but churn rates are higher. Users don’t complain about AI; they simply stop using it. This makes the problem invisible in standard feedback loops.
2. The lone user problem
The funniest moment in Samadder’s piece is when he refuses to eat ham alone in a bodega because a robot told him to. There’s a genuine insight here: AI recommendations work differently depending on whether you’re alone or with others. A group can laugh at a weird suggestion and pivot. A solo user following AI instructions can feel increasingly alienated from their own judgment. Consumer AI products rarely account for group dynamics, and the solo user experience — which is often the majority use case — suffers for it.
3. The “interesting versus enjoyable” conflation
AI systems trained on reviews, articles, and social media posts learn what people find interesting enough to write about. But writing-worthy and enjoyable are different things. The Bloomberg Mithraeum is fascinating and worth writing about. But visiting three underground attractions in a single day because the AI keeps selecting subterranean venues (postal tunnels, Roman temple, underground galleries) creates a monotonous emotional texture that no individual review would flag.
Implementation Note: If your AI recommends from a catalogue of items with reviews, add variety constraints that operate at the session level, not the item level. A user who loves three individual restaurants might hate being sent to three restaurants in a row. Sequence matters as much as selection.
4. The meta-awareness trap
Samadder’s most productive moment came when he used a second AI — his custom “RhikGPT” — to diagnose why the first AI’s plan wasn’t working. This meta-layer of AI reflecting on AI is increasingly common, but it creates a strange loop: users need AI to fix AI, which reinforces dependence on the very systems that are creating the problem. UK businesses should think carefully before adding “AI assistant for your AI assistant” features that compound complexity rather than resolving the underlying design failures.
The core opportunity
The Guardian experiment reveals something UK businesses should find encouraging: the bar for AI-powered consumer experiences is currently set at “individually good recommendations delivered without human understanding.” That bar is beatable.
The businesses that will win in consumer AI aren’t the ones with the best recommendation algorithms. They’re the ones that understand the full context of human experience — timing, energy, social needs, emotional arcs — and build AI systems that account for these factors rather than optimising around them.
Three factors that will define success in this space:
-
Model the human, not just their preferences. Preferences are inputs. Energy levels, social context, physical needs, and emotional states are the operating environment. Build for the environment, not just the inputs.
-
Design for deviation. The best moment in Samadder’s day was when he ignored the AI and went to a comedy night instead. AI systems that make deviation feel like failure (through guilt, lost progress, or broken sequences) will lose to systems that treat deviation as signal — “you went off-script here, so next time I’ll plan differently.”
-
Keep humans in the loop — literally. The comedy night worked because it involved other people. AI-guided experiences that maintain or create human connection will outperform those that replace it with algorithmic efficiency.
Strategic Reality: Consumer AI is not a recommendation problem. It’s a human experience design problem that happens to use recommendations as one input. UK businesses that understand this distinction have a genuine competitive advantage over platforms that treat better algorithms as the answer.
What to do this week
- Review your AI-powered features for the cocooning effect — are users becoming passive followers rather than active participants?
- Check whether your recommendation system accounts for physical and emotional context (timing, energy, social setting) or only preference matching
- Interview five recent users who completed an AI-guided experience but didn’t return — ask what was missing, not what was wrong
- Identify one place in your product where deliberate human friction (a staff recommendation, a personalised note, a phone call) could replace or augment an AI feature
This analysis is based on “I let AI guide me through London for a day. Why do I keep being sent underground?” by Rhik Samadder, published in The Guardian on 25 February 2026. Strategic analysis and business implications by Resultsense.