The Clique: Issue 004
23rd April 2026
“How can we tell if someone on the internet is human or an AI/bot?”
With ChatGPT Image2.0, the new human-verification step on Tinder and Zoom, and recent changes to Instagram, it turned out to be a good week to think about what that means. What infrastructure needs to be put in place to deal with it? What are the consequences if we don’t? Welcome back to the Clique.
This week:
One Thing I’ve Been Thinking About: authenticity in an age of convincing fakes
Three Stories: Claude’s biggest model upgrade yet, a new visual design tool for non-designers, and two major platforms asking users to scan their eyes
One Thing Worth Trying: ask AI to critique its own output
In Other News: Deep Research Max, OpenAI Codex’s biggest update yet, Adobe’s creative agent, Canva AI 2.0, and more…
Feel free to skip around.
1. One Thing I’ve Been Thinking About
The thing that gave creators their edge was always that being authentic at scale was hard. You had to actually show up, actually have a perspective, actually build trust over time. AI has started to collapse that barrier. Anyone can now produce content that sounds real and feels personal without the years of work or the genuine relationships behind it.
Adam Mosseri, who runs Instagram, wrote about this in a Threads memo late last year. He made a sharp observation: everything that made creators matter, the ability to be real, to connect, to have a voice that couldn’t be faked, is now suddenly accessible to anyone with the right tools. His platform-level answer is to surface more context about accounts so people can make more informed judgements about what they’re reading and who they’re reading it from.
That’s sensible, but it doesn’t really answer the harder question for individual creators and their audiences. If authenticity can be faked convincingly, what actually matters now?
My instinct is that it isn’t the content itself, but the context behind the content. The opinion that comes from real experience. The recommendation from someone whose track record you can check. The voice that has a consistent, verifiable history. The things that can’t just be generated from a prompt require actual human lived experience.
This doesn’t necessarily make the situation worse. It might just make the bar clearer. The people who were building genuine expertise and genuine relationships were always the ones worth listening to. The difference now is that the shortcut-artists are more visible, which might make the originals easier to spot, not harder. I’m not sure I’ve completely landed on a final answer yet, but it’s definitely interesting to think about.
2. Three Stories That Actually Matter
i. The most capable version of Claude yet is now available - and there’s one practical thing to check before you upgrade
The latest version of Claude is now generally available. Anthropic describes it as their strongest model to date (apart from Mythos), with particular improvements in reasoning, handling complex multi-step tasks, and reading images. It can now interpret dense diagrams, screenshots, and visual documents with significantly more accuracy than before.
Across the standard benchmarks used to compare frontier AI models, it leads GPT-5.4 Pro and Gemini 3.1 Pro on most tasks involving complex problem-solving, research, and document analysis. The gap isn’t enormous on every measure, but it is consistent. Early feedback from users is broadly positive; the model is more reliable on long, complicated tasks and less likely to give up on something difficult partway through.
There are some mixed early reactions too, though these are worth treating with caution given how soon they’ve come after launch. Some users report the model is more literal than before, meaning prompts that previously worked smoothly may need slight adjustment. These kinds of complaints are common in the first week after any major model update, and it remains to be seen how widespread they are in practice.
The most concrete factor for users to weigh is the effective cost of the latest architecture. While Anthropic’s headline pricing for Claude 4.7 remains at $5 per million input tokens and $25 per million output tokens, this specific model employs a new tokeniser that produces up to 35% more tokens from the same English text compared to the 4.6 and legacy series. Independent analysis has confirmed this "token inflation" effect, which effectively raises the cost of using the 4.7 model via API or in business workflows. Furthermore, this change impacts Claude.ai subscribers, as the higher token count drains message quotas more quickly than when using older models. For teams with Claude embedded in high-volume processes, it may be worth sticking with Sonnet 4.6 for routine tasks to maintain cost efficiency.
Sources: Anthropic · VentureBeat · Simon Willison
ii. A new tool from Anthropic turns a description into polished visual work without requiring any design skill
It’s been heralded as the ‘death of Figma’. Maybe that’s a bit of an exaggeration, but it’s definitely an impressive piece of kit.
Claude Design launched last week in research preview. It lets you describe your design, and it produces a fully structured first version that you can then refine through conversation, inline comments, or direct edits. Unlike general-purpose AI tools that output static images, it generates layered, editable work with a layout hierarchy and design structure built in. During onboarding, it reads your existing files to build a design system for your team, so every subsequent piece it produces uses your colours, typography, and components automatically.
Exports cover most of the formats knowledge workers actually use (Canva, PDF, PPTX, HTML) and for teams working with Claude Code, there is a direct handoff that packages the design into something a developer can build from straight away.
It is currently in research preview, available to Claude Pro, Max, Team, and Enterprise subscribers. It is powered by the Opus 4.7 model, which makes the timing of the two releases together fairly deliberate; you get the more capable model and a new tool that makes direct use of its improved vision.
Sources: Anthropic · TechCrunch · VentureBeat
iii. Tinder and Zoom are now offering iris-scan verification - an early sign of where the internet is heading
The idea that you might need to scan your eye to prove you are a real person on a consumer app would have seemed far-fetched a few years ago. It is now available on Tinder and Zoom.
Both platforms have partnered with World, a company co-founded by Sam Altman, to offer optional identity verification through iris scanning. The technology converts an eye scan into a unique cryptographic identifier. The system uses what is called a zero-knowledge proof: it can confirm that you are human without storing or transmitting the underlying biometric data, which means the verification is privacy-preserving by design. For people who prefer a lower-friction alternative, World also offers government ID scanning and selfie verification.
On Tinder, verified users receive a visible badge on their profile to signal that they are a real person rather than a bot or an AI-generated account. The rollout follows a successful pilot in Japan and is now global. On Zoom, the integration is aimed at professional meetings, where concerns about AI-generated participants and deepfakes in video calls have grown alongside the capabilities of the tools that create them. DocuSign has also joined the partnership.
The practical implication for most people is simply awareness: opt-in human verification is becoming a standard feature of platforms where trust matters. For anyone using Tinder or Zoom professionally or personally, the badge is now an available signal worth understanding. The broader question is what it says about the internet. We are starting to need third-party confirmation of our own humanity.
Sources: TechCrunch · Gizmodo
3. From the Blog
Few-shot vs zero-shot prompting: which approach and when - a practical guide to one of the most useful prompting decisions you can make, and how to know which approach fits your task.
Seamless content ingestion for a Claude-Obsidian second brain - a simple setup for quickly ingesting new content from the web, no matter the format. The more you can add to your second brain, the more it can work for you.
4. One Thing Worth Trying
Ask AI to check its own work
Most people never ask AI to critique what it just told them. This is worth changing.
After you get a response, be that an analysis, a written draft, or a recommendation, try sending one more message: “Now critique your answer above. What did you get wrong, overstate, or leave out? Then rewrite it with those issues corrected.” The model will often surface gaps it glossed over, qualifications it skipped, or assumptions it made without flagging them. The rewrite that follows is usually better than the original.
The technique is sometimes called LLM self-evaluation, and Anthropic built a version of this principle into Claude’s training as part of its Constitutional AI approach. But you don’t need to understand the mechanism for it to be useful. For anything high-stakes, where a confident-sounding but subtly wrong answer could cause real problems, this one extra step makes a genuine difference. It takes about ten seconds and costs nothing extra.
5. In Other News
And in no particular order…
Google has launched Deep Research and Deep Research Max, two autonomous research agents built on Gemini 3.1 Pro, capable of planning and executing multi-step research across the open web and private data sources simultaneously, generating inline charts and supporting MCP connections — the more powerful variant scores 93.3% on DeepSearchQA, up from 66.1% in December.
OpenAI’s Codex received its most substantial update since launch, adding background computer use, a built-in browser, persistent automations, memory, and more than 90 new plugins covering tools such as Jira, GitLab, and the Microsoft suite — used by more than three million developers weekly.
Canva AI 2.0 recasts the platform as a conversational, agentic design system, built on a proprietary design model that produces fully layered, editable output from a single prompt, with persistent memory that learns your brand style and connectors for Slack, Gmail, Google Drive, and Notion — currently available to the first million users to sign up via the Canva homepage.
Adobe launched its Firefly AI Assistant, a single conversational interface that can execute multi-step creative workflows across Photoshop, Premiere, Lightroom, Illustrator, and the rest of Creative Cloud — now in public beta.
OpenAI introduced GPT-Rosalind, a frontier reasoning model built for life sciences and drug discovery, supporting evidence synthesis, hypothesis generation, and experimental planning with access to more than 50 scientific tools and databases — working with partners including Moderna, Amgen, and the Allen Institute.
YouTube has rolled out a setting that allows users to fully remove Shorts from their feed by setting a daily time limit of zero minutes, available on iOS and Android via Settings > Time Watched > Shorts Feed limit. Maybe not an AI story, but definitely one I’m fond of.
Google added a new way to use AI Mode in Chrome, letting you open web pages side-by-side with AI Mode and search across your currently open tabs, images, and files for context-aware follow-up questions — currently available in the US.
Gemini 3.1 Flash TTS is Google’s new text-to-speech model, supporting multi-speaker dialogue, audio tags for fine-grained control over pace and delivery, and more than 70 languages — all audio watermarked with SynthID.
Tubi has become the first streaming service to launch a native app inside ChatGPT, letting users describe what they’re in the mood for in plain language and receive curated results across Tubi’s 300,000-plus titles.
GPT-5.4 Pro solved Erdős Problem #1196, a 60-year-old open conjecture in number theory, in approximately 80 minutes — using an approach that had not been tried in the nine decades since the problem was posed, and prompting Fields medallist Terence Tao to extend the proof into a new mathematical theory within 24 hours.
Claude Opus 4.7 has taken the top spot on the Vibe Code Benchmark with a score of 71%, ahead of GPT-5.4 at 67% — a benchmark that tests a model’s ability to build a fully working web application from a plain-English description, where no model scored above 25% when it launched four months ago.
OpenAI launched ChatGPT Images 2.0, its first image generation model with built-in reasoning, supporting 2K resolution, up to eight coherent images from a single prompt with consistent characters and objects across the batch, and improved rendering of non-Latin scripts. DALL-E 2 and DALL-E 3 will be retired on 12th May.
Google has added personalised image generation to the Gemini app, letting subscribers connect their Google Photos library so Gemini can generate images featuring specific people, pets, or places without needing manual uploads or lengthy descriptions.
Google released a native Gemini app for Mac, available free to all Gemini users on macOS Sequoia and above, with an Option+Space shortcut for instant access and the ability to share anything on screen directly with Gemini.
Chrome now lets you save your best AI prompts as one-click tools called Skills, which can be applied to whatever page you are viewing with a single click, with a ready-to-use Skills Library included for common tasks.
It has been an Anthropic-heavy week, which feels worth acknowledging. Two new products in the same window is unusual, and the combination is interesting. One raises the ceiling of what AI is capable of, while the other is designed specifically for people who don't have the technical background to use the first one to its full potential. More capable and more accessible at the same time is a combination worth paying attention to. I suspect we’re going to see more of this.
If you made it this far, I appreciate you!
Stay curious,
James
Enjoyed this issue? Consider forwarding to a friend or colleague!
Hey, look, there’s even a little button for it and everything 👇