Claude Is Fixing PS4 Games in shadPS4, and the Fixes Are Getting Merged

2026-10-01 · News and analysis

Affiliate disclosure: This guide contains affiliate links. As an Amazon Associate, we earn from qualifying purchases at no extra cost to you. Learn more.

Yesterday a fix landed in the PS4 emulator shadPS4 that made Uncharted: Drake's Fortune Remastered render its world again. Before it, gameplay was a grey void with floating particles and a HUD. The person who submitted it says, in the pull request itself, that Claude did the investigation and wrote the code change.

It is not the first one. It is the fourth AI-assisted fix to merge into shadPS4's mainline in under a month.

Three weeks ago, another contributor disclosed that Claude Opus 5 found a bug in the audio library and wrote both the fix and the tests. That merged too.

This is a bigger deal than any single patch, because in September we wrote that the emulators on your handheld ban AI code. mGBA closes AI pull requests on sight. SDL banned them in April. QEMU says it cannot accept the legal risk. That article ended by asking what happens to projects sitting between the ban camp and the decomp camp that had just used AI to finish Melee.

shadPS4 is the answer, and it has been writing it down since August.

The Short Version

• shadPS4 added a formal "A.I. Rules" section to its contributing guide on 5 August 2026. It does not ban AI. It requires disclosure, and says undisclosed use may get you banned from submitting.
• At least four AI-assisted fixes have merged into mainline since early September, the most recent on 1 October 2026.
• The 1 October one is the headline. Uncharted: Drake's Fortune Remastered rendered gameplay as a grey background with only particles and the HUD. The world is back.
• At least 17 pull requests now disclose Claude. That is a search floor, not a precise count, and several are still open.
• It is not Claude only. Some contributors disclose Claude Code alongside DeepSeek.
• A separate fork optimised The Last Guardian specifically, built by someone who states they have no coding knowledge and that Claude wrote all of it. Readback stalls dropped from roughly 350 per second to about six.
• The honest counterweight: the most thorough AI-assisted shadPS4 investigation of 2026 took four days, produced 15,668 lines of code, and contributed nothing upstream. The fix was one integer that already existed in the emulator.
• inFAMOUS: Second Son is not fixed. Its lighting bug is the same class as the one that investigation solved, which makes it the obvious next target. Nobody has done it.

What shadPS4's AI Rules Actually Say

The section is titled "A.I. Rules" and it sits in the contributing guide between the general style rules and the naming conventions. It was added on 5 August 2026. It is three bullet points, and they are worth reading in full because the shape of the policy is the story.

The first requires disclosure. AI use should always be disclosed and described, and a contributor found to have used AI without disclosing it may be banned from making further pull requests.

The second draws a line that most policies do not. Descriptions and comments should never be written by AI. Everything written for humans to read has to be written by a human.

The third puts the burden on the person, not the tool. Any AI-generated code a contributor wants to keep has to be pre-reviewed and polished, and the contributor must be able to logically explain their decisions, including the AI's, regardless of whether the code works.

That last clause is the important one. It is not "does the patch work." It is "can you explain why it is right." A maintainer can enforce that in review by asking a question, which is exactly what a ban cannot do.

This is roughly the Jellyfin position from our AI policy roundup: allowed with disclosure. What makes shadPS4 different from Jellyfin is that shadPS4 is a low-level emulator, which is the category everyone assumed would land on the mGBA side.

The Fixes That Have Actually Merged

Four changes disclosing Claude have merged into mainline since the start of September.

The Uncharted one is worth explaining, because it shows what this kind of work actually looks like.

In that game, gameplay rendered as a grey background with particles and the HUD and nothing else. No world. The cause was a redirect in the emulator that sends stencil reads to a depth target. That redirect was being followed for any image format, so when the game reused a memory location, a perfectly ordinary 960x540 colour target sitting at the stencil address got associated with a depth buffer it was not compatible with. Reads and storage writes to the colour target went through an invalid view. The submitter counted 750 of those views in a single session, 374 of them writes.

The fix stops a colour image being associated with stencil when a memory location is reused. Images that genuinely are stencil-compatible still get associated as before, so the games that depend on that path keep working.

There is a caveat the submitter states plainly, and it matters for the rest of this article. The patch alone does not bring the world back. You also need Readbacks Mode set to Precise. Neither one works without the other.

Hold onto that setting. It comes back.

The Fork Built by Someone Who Says They Cannot Code

Separately, on 29 September, a one-game performance fork of shadPS4 appeared for The Last Guardian.

The release notes are unusually honest. The creator says they wanted to see how far one game could be optimised with no coding knowledge, and that all code and measurements were done with Claude. They tested on that one game only, they flag the fork as unofficial and made for fun, and they explicitly ask people not to report bugs from it to the shadPS4 team.

The numbers they report are specific. Buffer readbacks that stall the game dropped from roughly 350 per second to about six. The GPU command thread moved onto a second CPU thread. Shader lookups got cheaper, and game threads now wait for their own readbacks instead of blocking the command thread. With the 60 FPS patch and without the 4K texture patch, they say 60 FPS is reachable in many areas depending on the PC, while the heaviest scenes with thousands of draws per frame still run lower.

Two things keep this in proportion.

First, the known issues section lists rare crashes seen once each and not reproduced. That is what a one-person, one-game optimisation pass looks like, and it is the reason the creator is pointing bug reports away from upstream.

Second, when we checked it had 44 downloads. This is a curiosity worth knowing about, not a distribution event. It is interesting because of who built it and how, not because of how many people are running it.

It does fit a pattern we have been tracking all year. Game-specific shadPS4 forks are how Bloodborne and Gravity Rush 2 reached PC in playable shape. The new part is that the barrier to making one has dropped to someone who does not write C++.

The Four-Day Hunt That Produced Nothing Upstream

Back in April, a developer spent four days using Claude Code to work out why DriveClub went black during night races. This is the most thoroughly documented AI-assisted emulator investigation we have seen, and it is also the strongest argument against getting carried away.

The symptom: at the start of night races in certain locations, the screen went pitch black for 10 to 30 seconds, with only the HUD visible, then mysteriously recovered.

The cause turned out to be a feedback loop between the GPU and the CPU. DriveClub does auto-exposure by having a compute shader write luminance histograms into a buffer, then having the CPU read that buffer, work out a target exposure, and write it back into a lighting buffer the fragment shaders read. Without proper synchronisation, the CPU was reading stale zeros. Its exposure integrator concluded the scene was dark and ramped its output: 2.59, then 7.84, then 24, then 90, then 179, then 255. At that point everything clipped to black.

The fix was one integer. Setting readbacks mode to Precise makes the emulator intercept page faults so GPU writes are synchronised to the CPU before the CPU reads them. That setting already existed in shadPS4. Nobody had publicly connected it to this game.

Here is the scale of the hunt behind that one integer: 31 numbered phases, 44 commits, 15,668 lines of code added, 54 investigation documents, 61 MB of accumulated logs, and 57 MB of shader dumps across 5,803 files. Almost all of the code was instrumentation.

None of it was contributed upstream. The developer said plainly that almost none of it was useful to upstream, because the diagnostic machinery was specific to one game. The fork stayed public as a study document.

Their conclusion is the thing worth quoting, because it is the opposite of the way this gets reported:

AI doesn't do everything. AI does what you ask it to. It has no domain initiative. It's not capable of telling you "this looks like a broken GPU to CPU readback feedback loop" unless you already know there's a thing called a readback.

They also noted that Claude suggested stopping several times, at points where the game had become playable and a workaround looked acceptable. The developer refused and kept going to root cause. Human stubbornness closed it, not model capability.

One more caveat on that fix. Precise readbacks cost roughly 12 FPS on AMD hardware and cause hangs in other games including Bloodborne. It is hidden behind advanced settings for a reason. It is not a free win.

Why This Splits From mGBA, SDL and QEMU

In our September piece on AI code bans we argued that decompilation can use AI because it has an oracle. You compile your C with the original compiler and compare bytes against the original binary. Either it matches or it does not, so wrong answers cost nothing. Emulator code has no such check, so a plausible but subtly wrong patch can pass review and bite someone two years later.

That argument still holds. shadPS4 has not found an oracle. What it has done is pick a different place to put the check.

mGBA, SDL and QEMU put the check at the gate. No AI code enters, so provenance stays clean and nobody has to evaluate whether a given patch is subtly wrong. The cost is that a tool which might genuinely speed up chronically understaffed volunteer work is ruled out entirely.

shadPS4 put the check on the contributor. Disclose it, be able to explain it, write your own prose. The cost is that it only works if maintainers actually enforce the explanation requirement in review, every time, under exactly the time pressure that made the ban attractive in the first place.

Two things make shadPS4's position more defensible than it looks from outside.

The licensing argument that drives SDL and QEMU cuts less deeply here. Those projects are foundational infrastructure shipped inside commercial products and distributed to hundreds of millions of devices. A provenance problem in SDL is a problem for every emulator above it. shadPS4 is a GPL application at the top of the stack, with a much smaller blast radius if something has to be rewritten.

And these fixes are unusually checkable. The Uncharted patch is not a vibe. It names a specific invalid view, counts 750 of them, explains which games use the same code path and confirms they still work. That is a reviewable claim. A maintainer can disagree with it on the merits. That is not true of every AI-written patch, which is exactly why the policy demands the explanation.

Worth noting for completeness: it is not only Claude. Some contributors disclose Claude Code together with DeepSeek. The policy is model-agnostic, which is the right way to write it.

Second Son Is the Next Obvious Target

inFAMOUS: Second Son has been running in shadPS4 for a while. It reached "ingame" status back in version 0.8.1, long before any of this, with no AI involved at all. It is broadly playable today.

It also has two open bugs that should look very familiar by now.

The first reports that the map and environment render almost completely black. Ambient and global lighting are missing or near zero. Street lamps and other local light sources still show up and light their immediate surroundings, which makes the rest look even darker. Enemies are hard to see.

The detail that matters: the lighting temporarily becomes correct in two situations. When the player takes heavy damage and the screen flash effect fires, and during certain small cutscenes. The reporter's own read is that the base scene lighting pass is broken while full-screen post-process effects bypass whatever is failing.

That is the DriveClub signature. A scene lighting path that has lost its exposure or its synchronisation, with post-process passes that do not depend on it still rendering correctly.

The second bug adds missing smoke and particle effects, broken dynamic lighting, and a log full of Vulkan errors saying Bc3UnormBlock 1D and Bc5UnormBlock 2D image formats are not supported. In that report, cutscenes render correctly and gameplay does not, which points at the gameplay rendering path specifically.

To be completely clear, because this has been circulating with the details wrong: Second Son's lighting is not fixed, and no AI-assisted work on it exists. We checked the pull requests and the issue tracker. Nobody has taken it on.

What makes it the obvious next target is that the map now exists. There is a documented investigation into this exact class of bug in this exact emulator, published in public, with the diagnostic approach written down across 54 documents. There is a disclosure policy that says you may use AI to do it. And there is a recent merged patch proving the maintainers will take a readback-adjacent rendering fix when the reasoning is laid out properly.

If Second Son gets fixed in the next few months, that is the chain that did it.

What This Means If You Just Want to Play

Practically, nothing to do today. There is nothing to install, no setting to change, and no new game that suddenly works on your handheld because of this.

If you run shadPS4 on a PC handheld, our shadPS4 setup guide is still the install path and the settings that matter, and the honest state of PS4 emulation on handhelds has not moved. It remains x86 only and enthusiast grade, with the ROG Ally X class of device as the realistic floor.

Two things are worth watching over a longer horizon.

Stability is the thing to track, not speed. The risk a ban is designed to prevent is a plausible patch that passes review and corrupts a save two years later. If shadPS4's approach is working, you will see it in the absence of weird regressions, which is hard to notice and easy to take for granted. If it is not working, you will see it as bugs that are hard to bisect.

The gate on PS4 emulation has always been contributor time. This is a volunteer project reverse-engineering a console. More people able to produce reviewable fixes is the single thing that moves the compatibility list, and compatibility is what decides whether PS4 emulation on a handheld is ever more than a hobby. Whether this particular accelerant is a good idea is genuinely still an open question. It is no longer a hypothetical one.

Frequently Asked Questions

Does shadPS4 allow AI-generated code?

Yes, with conditions. Its contributing guide added an "A.I. Rules" section on 5 August 2026 requiring that AI use always be disclosed and described, that descriptions and comments be written by humans, and that any AI-generated code be pre-reviewed and polished with the contributor able to logically explain every decision. Undisclosed AI use may result in a ban from submitting further changes.

Have AI-assisted fixes actually been merged into shadPS4?

Yes. At least four changes disclosing Claude merged into mainline between 6 September and 1 October 2026, covering the file system, the audio library, the shader recompiler and the texture cache. At least 17 pull requests in total disclose Claude, including ones still open.

What did the 1 October fix actually do?

It stopped a colour image being incorrectly associated with a stencil target when a memory location gets reused. In Uncharted: Drake's Fortune Remastered that bug meant gameplay rendered as a grey background with only particles and the HUD visible. The world renders again, though the game also needs Readbacks Mode set to Precise.

Did someone use Claude to get inFAMOUS: Second Son running?

No. Second Son has been running in shadPS4 since version 0.8.1, well before any of this, with no AI involved. It still has open bugs around dark environment lighting and missing particles, and as of 1 October 2026 no AI-assisted work on those bugs exists. Its lighting bug does look like the same class of problem an AI-assisted investigation solved in DriveClub, which is why it is the obvious next candidate.

Is this different from mGBA and SDL banning AI code?

Completely. mGBA closes AI-generated pull requests on sight and SDL banned them in April 2026, both primarily over licensing provenance. shadPS4 took the opposite approach, allowing AI with mandatory disclosure and a requirement that the contributor be able to explain the code. One puts the check at the gate, the other puts it on the person submitting.

What was the DriveClub investigation?

A developer spent four days in April 2026 using Claude Code to find why DriveClub went black during night races. The cause was a GPU to CPU auto-exposure feedback loop reading stale data and ramping exposure until the image clipped to black. The fix was a single setting that already existed in the emulator. The investigation itself ran to 31 phases and over 15,000 lines of mostly diagnostic code, and none of it was contributed upstream.

Does AI make emulator development faster?

Not on its own, and the developer behind the DriveClub work argues the opposite. Their conclusion was that AI has no domain initiative and cannot suggest a cause you do not already have the vocabulary for. What it did well was breadth: reading source in parallel, parsing 61 MB of logs, comparing decompiled binaries. The direction came from a human who refused to stop at a workaround.

Should I use the AI-built Last Guardian fork?

Only if you accept what it is. It is an unofficial one-game fork, the creator states they have no coding knowledge and that Claude wrote all of it, it was tested on that single game, and its own notes list rare crashes that were never reproduced. The creator specifically asks that bugs from it not be reported to the shadPS4 team. It had 44 downloads when we checked.

Will this break my emulator?

Nothing has broken so far, and these are normal reviewed changes in mainline rather than a separate experimental branch. The risk worth watching is the one the banning projects cite: a plausible patch that passes review and causes a subtle problem later. That is a slow signal, so judge it over months of release notes rather than any single update.

Related Reading

• AI Finished the Melee Decomp. The Emulators on Your Handheld Ban It.
• shadPS4 Setup Guide
• PS4 Emulation Status in 2026
• How Bloodborne and Gravity Rush 2 Reached PC: The shadPS4 Fork Era
• The Recomp Scene Split Three Ways Over AI
• Melee Is 100% Decompiled, and AI Finished the Last Stretch
• shadPS4 v0.17.0 Adds Game Compression
• ShadNet: shadPS4 Is Building Online Play Without PSN
• Two New PS5 Emulators Are Real

Written 1 October 2026. shadPS4's A.I. Rules wording and date, the merged changes and their disclosures, and the inFAMOUS: Second Son bug reports all come from the project's own public documentation, pull requests and issue tracker, read 1 October 2026. The Last Guardian fork details come from that fork's release notes. The DriveClub investigation figures and quote come from the developer's own published writeup from April 2026. This project merges changes most days, so treat the merged list as a snapshot. We have not run any of this software and we publish no frame rate figures of our own. All emulation should use game files and firmware from hardware you legally own.