Reverse Engineering & Beyond: The Most Creative Uses of AI in 2026

Mixed-media collage of an ancient scroll, protein ribbon, glowing point-cloud building, crystal lattice, code patterns, and audio waveform connected by a teal thread.

By Donato Technologies · Updated October 21, 2026 · 6-minute read

The short answer

What are the most creative uses of AI in 2026? Outside web development, seven stand out: reverse engineering, self-healing codebases, protein design, materials discovery, 3D scene capture, voice dubbing with lip sync, and reading unopened ancient scrolls.

The pattern in all seven: AI narrows the search. A person who knows the domain decides what’s real.

Why you should care: That pattern changes who you should hire. The scarce skill isn’t prompting. It’s building the check and judging the output.

Why this matters now

You’re already using AI for boilerplate, SQL, and refactors. That’s table stakes.

The more interesting work is happening in compiled binaries, research labs, and 2,000-year-old scrolls. We’re a staffing firm, so we read each one through a single lens: what does it ask of the people on your team?

Straight talk up front: some of this works today. Some is a promising demo. We’ll tell you which is which.

The seven at a glance

#	Use	What AI does	Where it stands
1	Reverse engineering	Explains and rewrites compiled code	Working, human-led
2	Self-healing repos	Finds bugs, fixes them, opens pull requests	In use, needs review
3	Protein design	Designs molecules that don’t exist in nature	Lab-proven, far from a drug
4	Materials discovery	Predicts new stable materials	Predictions at scale, few built
5	3D scene capture	Turns photos and video into 3D	Shipping, standards forming
6	Voice dubbing	Clones a voice, translates it, syncs the lips	Available, with fraud risk
7	Ancient scrolls	Helps experts read unopened, charred scrolls	First full scroll read, June 2026

1. AI that reads compiled code

What’s happening: Security teams plug AI into disassemblers like Ghidra and IDA Pro. The AI guesses function names, explains what a block of assembly does, and rewrites it as readable C.

Why you should care: Banks, utilities, and agencies still run software nobody has source code for. AI speeds up the slow first pass, so your scarce security people spend their time on judgment instead of tracing.

The catch: In a May 2026 test, an agent with full Ghidra access did real analysis on only 10–15% of functions in large binaries, until researchers gave it a way to measure progress. A 2026 benchmark on cryptographic binaries still had humans ahead. And malware can carry text built to hijack the AI that’s reading it.

2. AI that fixes its own repos

This goes well past autocomplete. Agent setups (Claude Code, LangGraph, n8n workflows) run tests in sandboxed containers, read stack traces, trace dependencies, and open pull requests.

Why you should care: Reviewing code is the bottleneck now. One 2026 study estimated agent adoption at roughly 22–29% of the 120,000+ GitHub repos it examined.

The catch: “Zero human intervention” sounds great until it hits production. In an early study of 567 agent-written pull requests, about 45% still needed human changes. For regulated or critical code, keep a person in the loop.

3. AI that designs proteins

Models like AlphaFold 3 predict how proteins fold. Newer tools like RFdiffusion go further and design new ones. A Nature paper in January 2026 described antibodies designed from scratch that bind chosen targets, confirmed with cryo-EM imaging.

Why you should care: Biotech runs on GPUs, data pipelines, and ML code. Someone has to build all of it. If you staff healthcare or pharma IT, this is where the ML engineer who speaks biology comes from.

The catch: The early designs bound weakly and needed tuning. The lab behind them says many steps remain before any of it becomes a therapy.

4. AI that finds new materials

Google DeepMind’s GNoME model predicted 2.2 million new crystal structures, about 380,000 of them stable. The targets include better batteries, solar cells, and semiconductors.

Why you should care: Energy, utilities, and manufacturing teams are watching this closely. The work is heavy on simulation and data, which means data engineers and high-performance computing skills.

The catch: At announcement in 2023, only 736 of those materials had been independently made in a lab. Predicted isn’t built.

5. AI that turns video into 3D

3D Gaussian Splatting, introduced by Inria researchers in 2023, turns photos or video into a photorealistic 3D scene you can move through in real time. No hand-built polygon models.

Why you should care: It makes digital twins of plants, buildings, and infrastructure cheaper to build. Bentley already ships it in its reality-capture software, and groups like Khronos are debating how to fit it into the glTF standard.

The catch: It’s a visual layer. It doesn’t replace the engineering data underneath, and the standards are still forming.

6. AI that dubs video in your voice

Models clone a speaker’s voice, translate it into other languages, and adjust the lip movements to match.

Why you should care: Training videos, support content, and internal comms can reach every language without a re-shoot. It also hits your hiring desk. The same tech lets someone fake a face and voice on a video interview.

The catch: Experian named deepfake job candidates a top-five fraud threat for 2026, and Gartner projects 1 in 4 candidate profiles could be fake by 2028. Remote IT roles with system access are the obvious target.

7. AI that reads burned scrolls

In June 2026, the Vesuvius Challenge announced the first complete reading of a Herculaneum scroll that was never unrolled. X-ray scans plus machine learning unwrap it digitally. Papyrus experts do the reading.

Why you should care: It’s the cleanest example of the pattern. The team describes the models as helping experts see the ink, not reading on their own. The project also found talent by running an open competition, and many people on the research team first showed up as contestants.

The catch: So far the text is Greek philosophy, not a lost masterpiece, and each scroll takes enormous effort. The next goal is reading 90% of four fully scanned scrolls within a year.

What this means for your next hire

The market is already moving:

U.S. AI-related job postings hit 80,597 in August 2026, up from 9,241 in January 2022 (Lightcast, via CFO.com).
40% of employers say they can’t find the AI talent they need, and 49% say their pay structures haven’t kept pace (Payscale).

What we’d do about it:

Hire for the feedback loop. The reverse-engineering agent only got useful once someone gave it a way to measure progress. Look for engineers who write the tests, benchmarks, and eval sets.
Hire strong reviewers. Review is the new bottleneck. In our view, a senior engineer who catches what an agent misses is worth more than another prompt wrangler.
Pair domain depth with engineering. Don’t hunt for unicorns. A biologist and an ML engineer who respect each other beat one résumé that claims both.
Test on a real problem. Vesuvius found people by letting them work on the problem in public. A small, timed work sample tells you more than a keyword match.
Verify identity once, live. Add one live identity check to remote IT hiring. It’s cheap compared to a fake hire with system access.

Federal, state, and prime contractor teams: PwC’s 2026 data, as summarized by Robert Walters, puts the AI pay premium as low as 16% in government and public sector roles, versus up to 118% in consumer markets. You won’t win on salary alone, so lead with mission, stability, and a clear growth path. And treat anything an AI reads, binaries included, as untrusted input. Check what your contract or agency rules say before AI tools touch controlled data.

Quick FAQ

Can AI reverse engineer software on its own?
Not reliably. It speeds up the first pass, but 2026 research shows agents skip most functions without a progress measure, and humans still lead on hard cases like cryptography.

Is it safe to let AI agents merge code without human review?
For low-risk repos with strong tests, some teams do it. For regulated or critical code, keep a human reviewer. An early study found about 45% of agent pull requests needed human changes.

Which skills are hardest to hire for right now?
Roles that pair deep domain knowledge with real engineering. Payscale reports 40% of employers can’t find the AI talent they need.

Where should a team start with non-web AI?
Pick a problem where you already have data and a way to check the answer. Reverse engineering for security teams and agent workflows backed by solid tests both fit.

Your turn

Which non-web AI project has your attention right now? Tell us. We’re curious what you’re seeing.

If you’re staffing a role that touches any of this, Donato places IT talent with U.S. companies. Send us the problem you’re trying to solve, and we’ll tell you what the market looks like.