Question 1

How does AI text detection work?

Accepted Answer

The text engine combines three complementary signals. Perplexity measures how predictable each token is given the context — LLM outputs tend to sit in a narrow, low-perplexity band because models favour likely words. Burstiness measures variance in sentence length and structure; human writing swings more than machine writing. N-gram fingerprints capture phrase patterns specific to each model family, including GPT-4o, Claude 3 and Gemini. The three signals are fused into a per-paragraph score, so you can see exactly which sections look human and which look generated, rather than getting a single verdict over the whole document.

Question 2

Which LLMs can you detect?

Accepted Answer

The text engine covers the widely deployed LLM families: OpenAI GPT-3.5, GPT-4, GPT-4o and GPT-4.1, Anthropic Claude 2, Claude 3 and Claude 3.5 Sonnet, Google Gemini 1.0 and 1.5, Meta Llama 2 and Llama 3, Mistral 7B and Mixtral, plus Microsoft Copilot (which wraps GPT-4). Our reference corpus is refreshed each time a major new model ships. Smaller fine-tunes derived from these base models inherit much of their parent's fingerprint, so they are typically detected alongside the base family even without dedicated training.

Question 3

How accurate is AI text detection compared to image detection?

Accepted Answer

Honest answer: text detection is harder than image detection. LLMs evolve faster than image generators and the statistical gap between human and machine writing narrows with every release. On unmodified GPT-4o prose our text engine sits around 85–90% accuracy at a reasonable false-positive rate. On heavily edited or human-reviewed AI text, it drops. Image detection on Midjourney or DALL-E 3 outputs is higher, above 94% on undegraded content. We return a calibrated score instead of a binary verdict precisely so you can weight each signal according to its known limits, and we always recommend text detection as decision support, never as a sole judge.

Question 4

Can it detect mixed content — partly human, partly AI?

Accepted Answer

Yes. This is one of the main reasons we return per-segment rather than per-document scores. For text, each paragraph gets its own confidence value, so you can see that paragraphs 3 to 7 look human while paragraphs 8 to 12 look LLM-generated. For images, the heatmap highlights which regions look synthetic, so partial inpainting and composited collages are flagged accurately. Mixed content is the realistic scenario in 2025: most workflows are human-authored but lightly AI-assisted, and our detector is tuned to respect that rather than forcing a black-or-white label.

Question 5

Is it useful for academic integrity?

Accepted Answer

Yes, with an important caveat. CHECKHC is designed as a decision-support tool, not as disciplinary evidence. A high AI score should trigger a conversation with the student, not an automatic sanction. Several universities worldwide have been sued successfully when they treated a detector verdict as proof. Our recommended workflow: run detection, review the per-segment highlights, discuss findings with the student, and only escalate if additional independent evidence supports the case. Used that way, the tool is extremely valuable for shaping honest, transparent conversations about AI assistance in coursework.

Question 6

What about AI-edited content vs fully AI-generated?

Accepted Answer

Editing with AI assistance (Grammarly, DeepL Write, light GPT rewording) typically registers as a moderate signal rather than a strong one, because the underlying human sentence structure remains visible. Fully AI-generated content registers strongly. Image inpainting and AI-assisted retouching of an otherwise human photo usually produces a partial flag on the affected regions, not the whole image. CHECKHC returns separate scores for full generation, partial generation and likely assistance, so you can set your own threshold for what matters in your workflow.

Question 7

Can I detect content from newer models like GPT-4o or Claude 3.5?

Accepted Answer

Yes. Our reference corpus is updated continuously rather than on a fixed annual cycle. GPT-4o (released May 2024) and Claude 3.5 Sonnet (released June 2024) have been supported since their first weeks of public availability. Subsequent minor releases inherit the parent model's fingerprints and are generally detected without an explicit retraining pass. The platform publishes a release-notes feed so customers can see exactly which model versions are currently covered and when each was added to the corpus.

Question 8

How does blockchain certification work for AI content detection?

Accepted Answer

After detection runs, you can optionally certify the result on-chain. CHECKHC computes a SHA-256 hash of your original content plus the detection report, anchors that hash on the Solana blockchain and stores a permanent copy on Arweave (200+ year archival window). The raw file never leaves your device; only the cryptographic fingerprint is written on-chain, which keeps the process compatible with GDPR. Anyone holding a copy of the original can then independently verify it is unchanged and was analysed at that exact timestamp. This is particularly valuable for newsroom archives, legal exhibits and editorial audit trails.

AI Content Detector

Why You Need a Multi-Format AI Content Detector

How Multi-Format AI Content Detection Works

Text Engine — Perplexity & Burstiness

Image Engine — Pixel Signatures

Video Engine — Temporal Coherence

Unified Report & Certification

Who Uses the AI Content Detector

University professor

Content editor

UGC platform moderator

Why CHECKHC Stands Alone

Verify Text, Images and Video Together

Frequently Asked Questions

Related Solutions