AI Tools That Feel Illegal to Use

Perplexity AI

Conversational search engine that cites sources inline. For research queries — synthesis of multiple documents, complex multi-step questions, building a working knowledge base fast — it genuinely beats Google’s AI Overviews. Wordstream’s July 2025 study found Google AI Overviews had a 26% error rate on PPC topics; Perplexity’s was 13%. Wordstream July 2025 study, cited in Skywork.ai comparison analysis. Independent publication, not vendor-funded; directional but not a controlled trial. For local search, shopping, or anything tied to Google’s ecosystem, Google still wins easily.

Reached 780 million monthly queries by May 2025 — up from 230 million in August 2024. That growth is organic and real.

⚠ Catch: Perplexity has faced credible allegations of scraping content from publishers that explicitly blocked it via robots.txt, then citing second-hand versions rather than original sources. For any content strategy work, that’s an attribution and ethical concern worth knowing. It also has no ads yet — but ads are coming, per the company’s funding requirements.

NotebookLM

Google’s document-grounded AI. You upload your own sources — PDFs, Google Docs, URLs — and it only draws from those. For synthesizing a body of research without hallucinating external facts, it’s unusually reliable because it literally can’t go off-script. The “Audio Overview” feature generates podcast-style discussions from documents, which is either useful or weird depending on your tolerance for synthetic conversation.

⚠ Catch: Hallucination risk when asked to connect ideas that aren’t explicitly stated in the uploaded documents. It can confabulate logical bridges. Always verify synthesized claims against the source material.

ChatGPT (GPT-4o)

The most versatile LLM available, still. Broadest application breadth — writing, coding, data analysis, image generation, voice input. Where it struggles: factual accuracy on anything recent without web search enabled, and a tendency toward confident-sounding plausible errors (hallucination). Pair it with Perplexity for any claim that needs to hold up under scrutiny.

⚠ Catch: Hallucination rate is meaningfully higher than Perplexity for research tasks. If you’re using it to draft content with cited statistics, every number needs independent verification. This is not optional; it’s been embarrassing in ways that are well-documented.

Claude (Anthropic)

Better than GPT-4o for tasks requiring long-form reasoning, document analysis, and careful, hedged output. The context window is among the largest available, which means it handles very long documents without losing track. Slower than ChatGPT at casual tasks; meaningfully better at tasks where being wrong has consequences.

⚠ Catch: More cautious than competitors, which reads as “safer” or “annoying” depending on your use case. For creative content that needs edge, it errs conservative. Not the tool for anything that needs to be edgy or provocative.

Gamma.app

AI-native presentation builder. Text prompt to full slide deck in under two minutes. For internal presentations, client proposals, and quick-turn deliverables, it removes the tedious formatting work that Canva and PowerPoint require. The aesthetic ceiling is lower than hand-crafted work — if visual design is a differentiator for you, Gamma won’t cut it. If it’s a vehicle for information, it’s excellent.

⚠ Catch: Output is recognizably AI-generated to anyone who uses it. For high-stakes client decks where design signals professionalism, this is a real limitation, not just aesthetic preference.

ElevenLabs

The best voice cloning and synthesis tool available, by a wide margin. For legitimate applications — branded voiceovers, multilingual content, audiobook narration, accessibility work — the output quality is genuinely stunning. 32+ languages. Voice cloning from as little as 1-5 minutes of audio.

The “feels illegal” sensation here is real: a voiceover that would have cost $500 and a week of scheduling takes eleven minutes.

⚠ Serious catch, not a minor one: In early 2025, a backlash erupted when users discovered ElevenLabs’ updated Terms of Service contained language granting the company a “perpetual, irrevocable, royalty-free, worldwide license” to use voice recordings and the models derived from them — even after account deletion. Kukarella documented this and publicly ended their ElevenLabs partnership, citing it explicitly. The ToS language is the actual text. Verify current terms before uploading personal voice data. ElevenLabs has since faced scrutiny on this. Additionally, NBC News documented that their voice cloning tools, like most competitors’ tools, have flimsy guardrails preventing nonconsensual impersonation. They’ve added verification steps since. Use for brand voice only — never clone another person’s voice without explicit written consent. The EU AI Act now requires AI-generated audio to be labeled as such.

Leonardo AI

Text-to-image generation with a generous free tier that genuinely edges out paid alternatives for most practical use cases. For marketing assets, product mockups, and concept illustration, the quality-to-cost ratio is hard to beat. More control over style consistency than DALL-E, which matters when you’re building branded assets.

⚠ Catch: Training data provenance is contested across the entire text-to-image industry. If your work involves clients in regulated industries (healthcare, legal, finance), check their IP policy before using AI-generated images commercially.

Uizard

Sketch-to-UI prototype in minutes. Draw a rough wireframe, upload it, get a working interactive mockup. For founders, product managers, and designers who need to communicate UI concepts quickly without burning engineering time on premature builds, this is the correct tool.

⚠ Catch: The output is a prototype, not production code. Teams have run into trouble treating Uizard exports as ready-to-ship — they’re not. The value is in the communication artifact, not the code.

Cursor

This is the one that genuinely changed how professional developers work in 2024–2025. VS Code fork, so zero switching cost. The agentic mode completes multi-file tasks — refactoring, feature builds, debugging across a whole codebase — in ways that GitHub Copilot inline completion doesn’t approach.

The academic data: a University of Chicago study on tens of thousands of actual Cursor users found 39% more merged PRs after Cursor’s agent became the default. Experienced developers showed higher adoption and better outcomes than junior developers — which is the counterintuitive part. The tools that make junior devs feel productive are often the ones that make experienced ones more so. Sarkar, U. of Chicago, Nov. 2025. Published by Cursor on their blog. Analyzed real usage data, not survey. Eligible organizations vs. baseline group. Directional on quality; PR merge rate as proxy for throughput.

Self-reported numbers go as high as 126% productivity improvement. That’s a survey figure. The real number depends heavily on your codebase, team practices, and how much you invest in learning the agentic features. Expect 20–40% in practice if you use it properly.

⚠ Catch: Memory-hungry (1–2GB more RAM than VS Code baseline; multi-agent mode can spike to 4GB+). On 8GB machines you will notice. The Pro tier ($20/month) is where the real value is — the free tier is limited. Also: a 2024 RCT found experienced open-source developers took 19% longer on some tasks with AI tools. Tool integration quality matters more than tool quality.

Sora (OpenAI)

Text-to-video generation at a quality level that was genuinely science fiction two years ago. For short-form content, concept visualization, and b-roll generation, it removes a category of production work that used to require a crew. Not there yet for broadcast-quality narrative content — the “Hollywood-level” claim is marketing.

⚠ Catch: Compute costs mean it’s expensive at scale. Deepfake and disinformation potential is the most serious ethical concern of any tool on this list — video is the hardest medium to detect as synthetic. OpenAI has implemented C2PA provenance metadata, which helps but doesn’t solve the distribution problem once video is shared outside controlled channels.

Runway ML

Video editing suite with AI features — background removal, inpainting, motion tracking, style transfer — that used to require After Effects expertise. For content creators, marketers, and small production teams, the capability jump vs. the learning curve is steep in a good direction. Better for editing existing footage than Sora’s generation approach; they solve different problems.

⚠ Catch: The free tier is genuinely limited. Quality at the free tier will frustrate you into upgrading or abandoning. Know that going in.

AgentGPT

Autonomous agent platform — you give it a goal, it breaks it into steps and executes them independently. Research tasks, content pipelines, data gathering. The “feels illegal” moment is watching an agent do a twelve-step research task while you do something else.

⚠ Catch: Reliability drops significantly on complex or ambiguous tasks. Agents can get into loops or produce convincingly formatted nonsense. Treat agent output as a first draft requiring human review, not a finished product. Don’t give it access to anything consequential — email sends, purchases, account changes — without a human in the loop.