Tell HN: Anthropic should make Claude the Author and me the Co-Author

The Hook

Today: Anthropic quietly changed how Claude Code credits its own work, and the developer internet noticed. The $20-a-month AI subscription battle just got its first honest 2026 scorecard. And Broadcom is projecting a 400% AI revenue surge over the next two years — the clearest forward dollar signal on the AI buildout we have seen this cycle.

The Cold Open

It is 11 p.m. A developer commits a pull request. The message reads: 'Co-Authored-By: Claude.' They hit send and close the laptop. But somewhere between Anthropic's servers and that commit log, a question got embedded that IP lawyers are only just starting to unpack — who actually wrote this code? Not in the sentimental sense. In the legal one. Tonight a Hacker News thread is pulling that question into the open, because Anthropic appears to have quietly changed how Claude Code attributes its own contributions. The rules of authorship in the AI era are being written in real time, one commit at a time. We are here for it.

The Signal

1. The Claude Code Authorship Dispute Is Now Public
Anthropic quietly changed how Claude Code handles attribution in code commits — and the developer community noticed. A Hacker News thread titled 'Anthropic should make Claude the Author and me the Co-Author' is surfacing a real tension: when Claude Code writes 80 percent of a function, the current convention labels the human as author and Claude as co-author. The community argues this is backwards — or at minimum, legally unclear. Open GitHub issues 66602 and 69835 document the behavior shifting without announcement. The IP stakes here are real. The US Copyright Office has ruled repeatedly that AI cannot hold copyright, but human attribution of AI-dominant work creates its own liability. When the human is listed as author of code they largely prompted into existence, they are making a legal claim about creative contribution that may not survive scrutiny. This is the friction point where the industry's most productive coding tool meets an unresolved legal framework. Practically: document your AI contributions clearly — not just for ethics, but for defensibility.

2. Google AI Pro vs. ChatGPT Plus — The 2026 Verdict
The first serious 2026 head-to-head of the two dominant $20-per-month AI subscriptions lands with a nuanced take: neither is an obvious winner. Google AI Pro earns marks for Deep Research, multimodal integration across Workspace, and Gemini's reasoning on long-context tasks. ChatGPT Plus holds ground on coding, image generation, and the breadth of GPT-4o's conversational range. The key differentiator in 2026 is workflow fit — Google wins if your working life runs in Docs, Sheets, and Gmail; ChatGPT wins if you are running autonomous tasks, custom GPTs, or need a coding copilot you can talk to like a colleague. The takeaway for readers: before you subscribe, map your top three actual AI use cases. The plan that fits your real workflow beats the one with the better marketing, every single time.

3. Designers — AI Is the Intern, Not the Executioner
Industry leaders speaking at London Design Fair pushed back hard on the AI replacement narrative this week. The consensus framing: generative AI is 'the intern in the office' — capable of strong first drafts, available at any hour, but lacking the judgment, taste, and client-reading skill that experienced designers carry. Agency heads cited using AI to expand what their teams can produce, not to cut headcount. The honest nuance the article buries: junior designers doing templated, low-judgment work are more exposed than senior strategists. The practical read for any creative professional — AI raises the floor of what can be produced cheaply, which means undifferentiated work gets commoditized. The answer is moving up the stack toward direction, synthesis, and taste, not retreating from the tools.

4. GuidedReview — A Human Checkpoint for AI-Generated Code
A new open-source tool called GuidedReview fills a gap that AI coding copilots quietly created: a structured review step before you sign your name to AI-generated code. Built by developer Nitish Arora and published to GitHub this week, it gives developers a guided, question-driven walkthrough of AI-written code — prompting reviewers to check for logic errors, security issues, and intent drift before committing. The accountability gap it targets is real: as AI writes more of the actual code, the diff-reviewed step in most teams' workflows has effectively stopped functioning — nobody deeply knows what the AI got right and what it hallucinated. GuidedReview is an early, practical answer to a problem that will only grow. Worth examining for any team that is shipping AI-generated code to production today.

5. Microsoft Project Zenith and F-Droid's AI Governance Moment
Two governance stories share space today. Microsoft announced Project Zenith, a new initiative within its AI infrastructure stack — details remain thin, but early signals point toward enterprise AI orchestration tooling. Meanwhile, F-Droid — the open-source Android app repository — announced it is modeling its generative AI usage policy on Debian's framework, one of the most thoughtful and structurally rigorous open-source governance models in existence. The F-Droid move is the more interesting story. Debian's approach distinguishes between AI as a build tool and AI as a component of distributed software, with specific disclosure requirements for each. If F-Droid's policy spreads as a template, it could define how open-source projects handle AI provenance and contribution disclosure for the next decade.

6. cognifity-verdict-inspect — Audit Your Own Chat History
A new Python package — cognifity-verdict-inspect — takes a novel approach to AI quality assurance: drop in your ChatGPT or Claude conversations.json export, and it runs a drift, hedging, and refusal analysis on your actual interaction history. It is essentially a quality scorecard for your AI, built from your real usage over time. The practical problem it solves is real: models hedge more, refuse more, or drift in tone in ways that are not obvious session to session but become visible in aggregate. This tool surfaces that drift so you can decide whether to adjust your prompting strategy, switch subscription tiers, or escalate feedback to the provider. Alpha-stage and early, but the concept is genuinely sharp and fills a gap no major AI provider has addressed.

7. Broadcom's 400% AI Revenue Projection
Broadcom is projecting a 400% surge in AI-related revenue over the next two years. Broadcom is a key infrastructure play — its custom AI chips, known as XPUs, power large-scale model training and inference for Google, Meta, and ByteDance. The 400% figure reflects analyst consensus on accelerating hyperscaler spend, not merely promotional guidance. The read for the broader AI landscape: the hardware buildout is not in a consolidation phase. It is re-accelerating. For practitioners, this means the hardware bottleneck on large-model inference is about to receive an extraordinary amount of capital, which has direct downstream effects on latency, availability, and cost curves for API-based AI services everyone in this audience uses daily.

8. Diffusion TV — Touching a Neural Network
Researchers published a paper (arxiv:2609.05404) describing Diffusion TV — a physical AI art installation that uses a modified CRT television as the interactive interface for a live diffusion model. Hardware modifications let users control the denoising process through physical dials on the chassis, making the model's latent space literally tangible to the hand. It reads like a stunt but the underlying research is serious: user studies showed that participants with no AI background who interacted with the physical installation developed significantly more accurate mental models of probabilistic image generation than control groups using a standard screen interface. Physical metaphors accelerate abstract concept formation. For anyone thinking about human-AI interaction design, this is a genuinely novel data point worth the read.

Quick Hits

• cognifity-verdict-inspect 0.1.0a16 — drop your ChatGPT or Claude conversations export in and get a drift-and-refusal audit back; alpha-stage but the concept is sharp enough to track.
• F-Droid adopts Debian AI governance framework — one of the most rigorous open-source AI policy structures just found a new home in the Android app ecosystem; watch for it to spread.
• Microsoft Project Zenith — enterprise AI orchestration tooling inbound from Microsoft; details sparse, but the name is in motion and the enterprise stack is consolidating fast.

The Anchor

Who Wrote This Code? The Claude Authorship Dispute Has No Clean Answer

When a developer commits a pull request and types 'Co-Authored-By: Claude,' they are making a legal claim, not just a social one. And right now, the rules governing that claim are being written in a Hacker News thread, a cluster of open GitHub issues, and the general productive chaos of an industry that moved faster than its legal frameworks.

The Hacker News post — 'Anthropic should make Claude the Author and me the Co-Author' — is doing something important: naming the inversion. When Claude Code writes 70 or 80 percent of a function, listing the human as author and the AI as co-author is not just philosophically imprecise. It may be legally misleading. The US Copyright Office has ruled repeatedly that AI cannot hold copyright — only humans or recognized legal entities can. But that ruling cuts both ways: if the AI did the majority of the generative work, the human attribution is a polite fiction with liability attached to it.

Anthropic appears to have been quietly iterating on this behavior. Open issues on the Claude Code GitHub repository — numbers 66602 and 69835 among others — document changes to how commits are attributed that were not announced, did not appear in release notes, and that developers only caught because the output changed between sessions. That kind of silent behavior drift erodes trust in a tool that is supposed to be a transparent collaborator, regardless of the legal question underneath it.

The deeper problem is structural. Authorship conventions in software development evolved entirely for human collaboration. When two engineers pair-program, the commit log captures both names. When one writes and one reviews, conventions exist for that too. None of those conventions were designed for the scenario where one party generates 80 percent of the code in 0.3 seconds and has no legal personhood, no accountability, and no stake in the outcome.

What should developers do right now? Three concrete things. First, maintain a session log — note what you prompted, what Claude generated verbatim, and what you modified. A simple gitignored markdown file per session is enough. Second, adopt an honest commit-message convention: 'Primary author: Claude Code — reviewed, refactored, and tested by [name]' is more defensible than the inverse and still credits the human contribution accurately. Third, watch for Anthropic's official guidance. They will be forced to publish a clear policy by the weight of enterprise customer demand alone, and when they do, it will set industry norms quickly.

This dispute is small today — a thread on Hacker News, a handful of GitHub issues. But the underlying question of who legally authors AI-assisted work is the live IP dispute of 2026. Every tool in the AI coding stack is going to face a version of it. The answer matters for contracts, open-source licensing terms, and enterprise liability in ways that are still unresolved. The friction started here, in public, in September. The resolution will take years and probably a court case or two. Start your session logs now.

Deep Dive

Diffusion TV — What Happens When You Make a Neural Network Tangible

At its core, a diffusion model works by learning to reverse a destruction process. During training: take an image, add Gaussian noise incrementally across hundreds of discrete timesteps until you have pure random static. The model learns the reverse trajectory — starting from noise, predicting what noise was added at each step, and removing it progressively until a coherent image re-emerges. At inference time, you start from random noise, condition the model on a text prompt or other signal, and the denoising process runs forward through those timesteps to generate your output. The image is not retrieved or assembled — it is grown out of noise under the guidance of your conditioning signal.

The Diffusion TV installation (arxiv:2609.05404) takes this mechanism and makes it physically interactive. The researchers modified a CRT television — the kind with an analog electron gun and a phosphor screen — to serve as the live output display for a diffusion model running inference in real time. Physical controls mounted on the TV chassis map directly to model parameters: one dial adjusts the number of denoising timesteps (fewer steps means a coarser, noisier result; more steps means finer resolution and coherence), another controls the classifier-free guidance scale (how strongly the prompt pulls the output versus free generation), and a third injects deliberate noise mid-inference, visibly destabilizing the image before it re-resolves.

The choice of a CRT is not incidental. The phosphor persistence and scan-line aesthetics of a CRT screen create a perceptual bridge between the noisy early timesteps of a diffusion run and the physical experience of tuning an analog television — static resolving into signal as you turn the dial. That perceptual metaphor is not decorative. It is doing cognitive work. The researchers hypothesized that users with prior experience of analog static-to-signal would use that embodied memory as a scaffold for understanding what the model is doing computationally.

The user study results support this. Participants with no AI or ML background who interacted with Diffusion TV showed significantly more accurate mental models of probabilistic image generation compared to control groups who used a standard browser-based interface with identical underlying functionality — same model, same parameters, different form factor. The embodied interaction group could more accurately describe what 'more denoising steps' meant for output quality, what happened when guidance scale was reduced, and why the model sometimes produced unexpected outputs.

The mechanism the paper proposes: embodied action creates causal model formation. When your hand controls the noise level and your eye watches the image resolve step by step, you are not observing an output — you are constructing a causal model of the process. The dominant paradigm of text-prompt-in, image-out hides all of the model's interesting computational structure from the user. Interfaces that expose that structure — even approximately, even through metaphor — may produce users who are fundamentally better equipped to prompt, debug, and reason about generative AI outputs.

The limitation is obvious: CRT televisions are not a scalable consumer interface. But the principle is transferable. What would a diffusion model's latent space controls look like as a standard UI component in a web app? What does 'expose the mechanism' mean at scale? Those questions are genuinely open. This paper is a rigorous early sketch of why they matter.

One Technique

The Authorship Audit — Know What You Actually Wrote

Before your next AI-assisted coding session, set up a lightweight authorship log. Create a file called AI-SESSION.md in your project root and add it to your .gitignore. For each significant block of code you generate with Claude Code or a copilot, note three things: the prompt you used, what the model returned, and what you actually kept or modified. At commit time, write a message that reflects the real ratio — something like 'Implemented auth middleware: Claude generated the initial scaffold, rewrote the token validation logic and error handling.' This takes two minutes per session and gives you defensible documentation of your creative contribution. As the authorship dispute matures legally, teams with session logs will be in a categorically stronger position than teams without them.

One Prompt

Use this prompt to run a structured pre-commit review of any AI-generated code before you sign your name to it:

You are a senior code reviewer. I am going to give you a code block that was generated by an AI assistant. Your job is to review it as if you were fully responsible for it in production.

For each section of the code, tell me:
1. What this code does in plain English (not restating the code itself)
2. Any logic errors or edge cases it misses
3. Any security concerns — injection, auth bypass, missing input validation
4. Any places where the AI may have hallucinated a method, library, or API behavior that does not exist
5. Your overall confidence rating (1 to 5) that this code is production-safe

Code to review:
[PASTE YOUR AI-GENERATED CODE HERE]

Be direct. Flag anything you would push back on in a real code review. Do not soften findings.

One Tip

Turn on inline Git blame for every AI-generated file. In VS Code, install GitLens and enable inline blame view. When you are debugging AI-generated code weeks after the session, you want to know immediately whether a line came from a human decision or an AI generation run. This requires no new workflow — just a one-time settings change. It turns your git history into an honest record of contribution over time, which matters more the larger the fraction of your codebase that AI writes.

Tool of the Day

GuidedReview — github.com/nshntarora/guidedreview

An open-source tool that gives developers a structured, question-driven walkthrough of AI-generated code before committing. It prompts reviewers to check for logic errors, security issues, and intent drift — the exact things that slip through when you are reviewing a diff that you did not mentally construct yourself. Genuinely useful for any team where AI is generating more than 20 percent of production code. Honest limit: it is a guided checklist, not a replacement for deep human judgment — treat it as the floor of your review process, not the ceiling. Works best deployed as a pre-commit hook so the checkpoint cannot be skipped under deadline pressure.

Signature Bites

• The authorship inversion: When Claude writes 80 percent of a function, listing the human as sole author is not just imprecise — it is potentially a legally exposed claim.
• The $20 question: Google AI Pro wins on Workspace integration; ChatGPT Plus wins on autonomous tasks and coding copilot range. Map your use case before you subscribe.
• The intern frame: 'AI as the intern in the office' is the creative industry's healthiest reframe of 2026 — accurate, actionable, and blessedly non-hysterical.
• 400% in 24 months: Broadcom's AI revenue projection is the clearest infrastructure acceleration signal of the quarter — hardware spend is re-accelerating, not plateauing.

Joke of the Day

A developer commits a PR. Commit message: 'Co-Authored-By: Claude.' The IP lawyer asks: 'So Claude wrote it. Who owns it?' Developer: 'Me.' Lawyer: 'Who reviewed it?' Developer: '...Claude.' Lawyer: 'Who approved the architecture?' Developer: '...Claude.' The lawyer closes the laptop. 'I'll see myself out.'

Fact of the Day

The US Copyright Office has officially ruled — in 2023 and again in a more detailed 2024 guidance — that AI-generated content is not eligible for copyright protection. The 2024 ruling specifically addressed AI-assisted works, concluding that copyright attaches only to the human-authored elements of a work, not to the portions generated by an AI system. This means the 'human as primary author' convention currently common in AI coding tools may create the legal impression of human authorship over content that, under current US law, does not qualify for copyright protection in its AI-generated portions.

Stat That Matters

400% — Broadcom's projected AI revenue growth over the next 24 months. Broadcom supplies custom AI accelerator chips to Google, Meta, and ByteDance for large-scale model training and inference. A 400% revenue projection from a company embedded at the foundation of hyperscaler AI infrastructure is not a marketing claim — it reflects contracted purchase orders and multi-year capacity commitments from the largest AI operators on the planet. The implication for practitioners: the hardware layer of AI is in an acceleration phase, not consolidation. API latency will keep falling. Inference cost curves will keep improving. The buildout is ahead of us, not behind us.

Trends

Agentic AI is the dominant lane in today's corpus — 146 stories tracked, nearly double any other category. The signal is simple: autonomous, tool-using AI is no longer a trend category. It is the baseline that every other story assumes. Funding is surging in parallel at 77 stories, tracking directly with the infrastructure acceleration that Broadcom's projections confirm at the hardware layer. The convergence of agentic deployment scale, capital acceleration, and now active public IP disputes around AI authorship suggests Q4 2026 will be the quarter where enterprise AI governance transitions from best practice to legal and contractual requirement. The three lines are converging faster than most governance frameworks are moving.

Bold Prediction

Anthropic will publish an explicit, public AI authorship attribution policy for Claude Code before December 31, 2026 — one that defines the recommended commit-message convention, directly addresses the copyright ambiguity for AI-assisted work, and likely introduces a machine-readable metadata format for marking AI contribution ratios in commit history. The Hacker News pressure is real. The open GitHub issues are accumulating. And the legal exposure to enterprise customers using Claude Code without clear attribution guidance is too significant for Anthropic to leave unaddressed heading into 2027 renewal cycles. Call: official policy published by year-end 2026. Hold this to account.

Paper Watch

Diffusion TV: Experiencing Diffusion Models through Tangible, Embodied Interaction — arxiv:2609.05404

Researchers built a physical AI art installation using a modified CRT television as an interactive interface for a real-time diffusion model. Users control the denoising process — the core computational step of image generation — through physical dials on the TV chassis. Key finding: participants with no AI background who used the physical installation developed significantly more accurate mental models of probabilistic image generation compared to control groups using a standard screen-based interface with identical functionality. The paper's argument: 'type prompt, get output' interfaces systematically hide the interesting and mechanistically useful structure of generative models from users, and embodied interaction closes that gap. Practical implication for builders: AI interfaces that expose their computational mechanism — even approximately, even through metaphor — may produce fundamentally more capable users than black-box interfaces do.

Founder Spotlight

Nitish Arora — GuidedReview

Nitish Arora shipped GuidedReview this week: an open-source tool that inserts a structured human checkpoint between AI-generated code and a git commit. The strategic read: Arora identified the exact accountability gap that every major AI coding copilot created and none of them addressed — the moment where a developer is about to commit code they did not mentally write themselves, and needs a genuine forcing function to actually review it rather than skim it. This is a founder move worth tracking because the problem it solves scales directly and proportionally with AI adoption. The more AI code in production codebases, the more this checkpoint matters. Early, open-source, and positioned in a gap the incumbents left open — textbook early market positioning in an expanding problem space.

Quote

'Firms are more likely to use [generative AI] as the intern in the office than as a replacement for skilled staff.'

— Industry leaders at London Design Fair, as reported by The Guardian, September 7, 2026.

Learner's Edge

What Is Latent Space?

Every generative AI model — image, text, audio — works by compressing its training data into a high-dimensional mathematical structure called a latent space. Think of it as a vast map where similar concepts cluster together: 'dog' and 'puppy' are close; 'dog' and 'skyscraper' are far apart. The model has learned the geometry of that map from billions of examples.

When you give the model a prompt, it locates a region of that map that corresponds to what you described, then navigates from there to produce output. In a diffusion model specifically, the navigation happens step by step — starting from a random point in the map (noise), progressively moving toward the region that matches your conditioning signal (the prompt), removing noise at each step until something coherent forms.

Understanding latent space unlocks several mysteries about model behavior. Why does changing one word in a prompt sometimes produce a wildly different image? You moved to a different region of the map. Why can AI blend concepts — a dog that looks like a Renaissance painting? Regions of the map can be interpolated. Why do hallucinations happen? The model reached a region that is internally coherent within its geometry but does not correspond to anything real. The latent space is the model's entire world. Everything it knows, and everything it can generate, exists there.

Sign-off

That is THE AGENT SIGNAL for September 7, 2026. Tomorrow we are watching whether Anthropic responds — formally or informally — to the authorship attribution pressure that is now publicly visible and accumulating. The quiet change is in the open. The next move is theirs.

Sources

1. Tell HN: Anthropic should make Claude the Author and me the Co-Author — news.ycombinator.com
2. Google AI Pro vs. ChatGPT Plus: Which Is Better in 2026? — provenbrief.com
3. Designers should not fear being replaced by AI, industry leaders say — theguardian.com
4. Show HN: GuidedReview – Review AI-generated code before you sign your name to it — github.com
5. PI A.M. Edition: Microsoft announces Project Zenith, F-Droid plans to model generative AI usage policy after Debian — sspai.com
6. cognifity-verdict-inspect 0.1.0a16 — pypi.org
7. With AI Revenue Set to Surge 400% Over the Next 2 Years, Broadcom Stock Looks Like a Buy on Recent Dip — Motley Fool
8. Diffusion TV: Experiencing Diffusion Models through Tangible, Embodied Interaction — arxiv.org