Is AI actually used in data recovery — or is it mostly marketing?

The honest answer is “yes, in parts”

The phrase AI data recovery sounds as if a model can look at a dead drive, understand what the owner wants and bring the files back. That is not how a real case breaks down.

There are at least four separate jobs:

1. stabilise the storage device;
2. acquire as much source data as possible;
3. reconstruct file systems, volumes and files;
4. validate, repair and present the result.

AI is already credible in parts of jobs 3 and 4. It may also help an engineer make decisions during triage. It does not replace the physics in jobs 1 and 2. A failed read/write head still needs a safe mechanical route. A controller still has to return data. A missing encryption key does not become less missing because a model is confident.

That distinction is the thread running through every claim in this article.

Which companies openly say they use AI?

As checked on 27 July 2026, at least two well-known consumer recovery-software brands make explicit public claims.

Wondershare Recoverit describes its product as an AI-powered data recovery solution. Its current material applies the term to device and scenario identification, scanning, media-file reconstruction and photo or video repair after recovery.

EaseUS Data Recovery Wizard labels file repair as AI-powered and says damaged MP4, MOV and JPEG files can be repaired during recovery. EaseUS also uses the AI label in separate media-repair products.

These are official claims, so it would be wrong to say that nobody in data recovery admits using AI. It would be equally wrong to treat every feature under that banner as the same technical process. Identifying a camera model, choosing a scan plan, matching video fragments and synthesising missing picture detail are four different tasks.

The vendors’ performance percentages are marketing claims made in their own test conditions. We have not used them as evidence of what will happen in an individual case.

Why do professional laboratory suppliers sound different?

The public materials we reviewed from two established professional tool suppliers use a different vocabulary.

ACE Lab describes PC-3000 in terms of direct device communication, firmware access, diagnosis, imaging, supported controllers, file systems and RAID reconstruction. DeepSpar’s Disk Imager brief describes head-by-head handling, multiple imaging passes, timeouts, resets and different read commands for different problem areas.

Neither of those reviewed product pages makes an AI claim. That observation is narrow and deliberate:

• it does not prove that the companies never use machine learning;
• it does not reveal what a laboratory may have built internally;
• it does show what those suppliers consider important enough to explain publicly.

In physical recovery, deterministic control is valuable. An engineer needs to know which command was sent, how a drive responded, which sectors were read and what changed between attempts. A fashionable label is not a substitute for that audit trail.

Where AI could genuinely help

1. Recognising file fragments

When file-system metadata is damaged or missing, recovery may rely on file carving: identifying data by its internal structure rather than by a directory entry. The hard version begins when fragments have no useful header, sit out of sequence or could belong to several formats.

This is a real machine-learning research problem. The FiFTy research project trained a neural network to identify 75 file-fragment types and reported that it outperformed the comparison tools on its test dataset. The remaining classification errors are a reminder that a prediction is not certainty.

In a recovery workflow, a classifier could rank candidate fragments or narrow the search. A validator would still need to test whether the assembled file is structurally coherent and whether it actually opens.

2. Reassembling damaged video

Modern cameras and phones can produce large, fragmented video files. If recording stops abruptly, the media may contain useful video and audio data without a complete container structure.

A model could help identify which fragments share encoding characteristics, estimate their likely order or flag a discontinuity. This is close to some of the use cases consumer vendors already market.

There is an important evidential line. Rebuilding a container around surviving original frames is recovery. Generating plausible frames to hide a gap is restoration. The second result may be visually pleasing, but it must never be presented as the original recording.

3. Prioritising a failing image

Imaging a weak drive is a resource-allocation problem. Some areas read cleanly; others cause long timeouts or destabilise a particular head. Existing professional imagers already let engineers use maps, passes, time limits and file-system knowledge to choose the order of work.

A future system could learn from error patterns and recommend:

• which head or address range to read first;
• when retries are producing diminishing returns;
• which files depend on the weakest areas;
• whether a changed power, timing or read strategy improved the result.

The safe design is decision support. The engineer approves the plan, limits the attempt and can explain why it was chosen. An opaque system repeatedly stressing a failing drive because its score might improve would be dangerous automation.

4. Finding anomalies in large recoveries

A recovered dataset can contain millions of objects. Machine learning could help cluster similar files, detect damaged outliers, identify duplicates or flag items whose extension conflicts with their content.

For a customer who needs one project folder urgently, assisted search could also reduce the time between raw recovery and a useful file list. Again, the model would be finding candidates, not deciding that an absent file never existed.

5. Learning from case notes without exposing case data

Laboratories accumulate diagnostic knowledge: controller behaviour, common firmware symptoms, compatible donor details and the consequences of particular interventions. A tightly controlled retrieval system could help an engineer find relevant internal notes faster.

That use is less dramatic than a robot repairing a drive, but probably more valuable. It also raises a hard governance question: can the system learn from technical patterns without ingesting customer filenames, documents or personal information?

What AI cannot do

AI works with information that exists in its input or with patterns learned elsewhere. It cannot recreate missing historical truth.

This is why “AI restored the photo” can mean two very different things. It may mean that original encoded data was found and made readable. It may mean that a generative model created a believable face or background. Only the first is data recovery.

Privacy is not a side issue

A drive sent for recovery may contain passports, health records, legal files, children’s photographs, source code, payroll information or client databases. Feeding that material into an online AI tool because it is convenient would create a new disclosure risk.

The UK Information Commissioner’s Office guidance applies data-protection principles including purpose limitation, data minimisation, accuracy, storage limitation, security and accountability to AI systems processing personal data.

For a recovery laboratory, that points towards strict controls:

• process locally or in an explicitly approved private environment;
• send only the minimum data needed for a defined task;
• prevent customer content from becoming training data;
• record the model, version, input and output where the result matters;
• keep a human reviewer responsible for the conclusion;
• delete intermediate material under a documented retention rule.

NIST’s AI Risk Management Framework uses the functions govern, map, measure and manage. Those ideas fit this field well: define the task, understand the risk, test the output and retain control.

What would convince us that an AI recovery feature is useful?

Not a glowing interface and not the word intelligent.

We would want to know:

1. What exact decision does the model make?
2. What data was it trained and tested on?
3. How does it perform on devices and file types outside that dataset?
4. Does it preserve the original evidence?
5. Can a technician inspect and override the result?
6. What happens to customer data?
7. How are false positives labelled?
8. Can another expert reproduce the result?

FiFTy is a good example of useful honesty. It identifies a bounded problem, publishes a dataset and reports an error rate. “Recovers anything with AI” tells a customer almost nothing.

Our present position

This article is not an announcement that Datraction uses generative AI on customer cases. It is a map of where the technology is publicly claimed, where research shows promise and where the physical or cryptographic limit remains unchanged.

AI can become a strong assistant in classification, triage, validation and knowledge retrieval. The standard should be the same as for any laboratory tool: it must improve the recoverable result without taking uncontrolled risks with the only copy.

The future of data recovery is likely to include more machine learning. The discipline will still depend on something older and less marketable: preserve the source, separate fact from inference and never manufacture certainty where the evidence has run out.