Every day, thousands of people ask ChatGPT, Claude, or Gemini to "make me a PowerPoint" — and the AI hands back code: a pptxgenjs script, a python-pptx program, sometimes a VBA macro. The user doesn't want code. They want the deck.
We operate the browser tools those users land on next (JavaScript to PPT, python-pptx online, Code to Slides) — which puts us in an unusual position: we may hold the largest real-world corpus of AI-written presentation code that actually gets executed. This report shares what's in it.
The dataset
- 27,600 script executions recorded between June 15 and July 31, 2026 (~600/day)
- 16,315 pasted code submissions on the code-to-PowerPoint tools
- Full production data for the period — no sampling
How much of it is AI-written? We can prove a floor: 42% of pasted scripts name an AI assistant in their own code or comments — 34% mention Claude/Anthropic, 2.4% ChatGPT/OpenAI, 3.2% Gemini/DeepSeek/Copilot/Grok, and 2.1% carry a generic "generated by AI" header. The real share is higher: most AI output has no marker at all, and a manual review of submissions found the overwhelming majority follow unmistakable AI patterns (uniform comment style, placeholder image URLs, "Slide 1/2/3" scaffolding).
Finding 1: AI deck code mostly works — when something runs it
Overall, 85.1% of executions succeed. On the main code tools:
| Tool | Runs | Success rate |
|---|---|---|
| JavaScript (pptxgenjs) → PPT | 6,208 | 92.5% |
| HTML → PPTX | 2,819 | 93.6% |
| Code to Slides (multi-language) | 3,931 | 83.7% |
| python-pptx → PPT | 1,056 | 79.3% |
Two caveats make these numbers more interesting, not less. First, they are post-repair: our sandbox auto-fixes a class of AI mistakes before execution (more below), so the raw quality of AI code is lower than the success rate suggests. Second, python-pptx code fails meaningfully more often than pptxgenjs code — Python deck scripts are longer, more stateful, and give the model more API surface to hallucinate.
Finding 2: the failure taxonomy
Setting aside account and quota walls, the real technical failures rank like this:
| Failure | Count | What it is |
|---|---|---|
| Not a deck generator | 372 | Code that runs but never creates a presentation — the AI produced analysis code, a web page, or a fragment |
| python-pptx runtime errors | 221 | Wrong argument types, invented methods, out-of-range indexes |
| Syntax errors | 156 | The script cannot be parsed at all — truncated AI output is a common cause |
| Wrong-runtime pastes | 272 | Python pasted into the JS tool (107), VBA macros (100), React components into non-React tools (65) |
| HTML paste with no convertible content | 88 | The pasted HTML contained nothing renderable as slides |
The VBA number deserves a highlight: 100 people in six weeks pasted a ChatGPT-generated VBA macro — code that requires desktop PowerPoint on Windows — into a browser tool, because that's what the AI gave them. AI assistants still routinely reach for 2010-era automation answers.
Finding 3: hallucinated APIs are real, but rarer than you'd think
We scanned all python-pptx submissions that use shape enums (1,380 scripts): 10 referenced MSO_SHAPE members that don't exist — inventions like MSO_SHAPE.TEARDROP (the real member is TEAR). That's under 1% — but each one is a guaranteed crash with a baffling AttributeError for a non-programmer. Our runner now auto-maps invented enum names to their nearest real member before execution, which is exactly the kind of repair a human developer would never need and an AI-era runtime cannot skip.
The bigger repair category is mundane: missing writeFile() calls, await outside async functions, image URLs that 404 (AI models love inventing https://example.com/chart.png). Our pre-fixer ladder addressed enough of these to lift the JS success rate by roughly five points.
Finding 4: these scripts are big
The median pasted script is 26,131 characters (~700 lines); the 90th percentile is 54,754 characters, and the largest submission was 1.09 MB of generated code. AI assistants do not write minimal examples — they write the whole deck, one hardcoded string at a time. That's also why "just fix it yourself" isn't realistic advice for the people pasting it.
What this means
- "AI can't write working code" is outdated — for this domain, >9 in 10 pptxgenjs scripts execute after light repair.
- The gap has moved from generation to execution. The AI writes the deck script in seconds; the user's actual problem is having no runtime. That's a tooling gap, not a model gap.
- Model vendors could close obvious holes: stop emitting VBA by default, stop inventing image URLs, and truncation mid-script remains the most user-hostile failure.
Methodology & reuse
Counts are aggregate SQL over our production execution log and submission archive (June 15 – July 31, 2026); no script contents beyond automated pattern counts were reviewed for this report, and no user data is included. You are welcome to cite this report with attribution to Sharayeh (link this page). We plan to refresh the numbers quarterly — the dataset grows by roughly 18,000 executions a month.
Want to see the runtime these numbers come from? Paste any AI-generated deck script into JavaScript to PPT or python-pptx online — try it without sign-up.