1. The Question Every AI Team Is Asking in 2026
Synthetic data pipelines have crossed a threshold. GPT-4-generated datasets, the Self-Instruct method, the Alpaca instruction-following dataset at $600, Phi-1.5 and Phi-2 trained primarily on synthetic "textbooks," GPT-4-generated preference pairs used in early alignment work — the empirical record now shows synthetic data producing competitive results on benchmarks that previously required large-scale human annotation pipelines.
This has created a genuine decision point, not a philosophical one. ML engineers planning the next training run are asking a specific question: given our task, our domain, and our quality requirements, is the cost of human expert annotation justified over a synthetic data pipeline? And they deserve a specific answer.
The answer is not "it depends." The answer is a framework with clear conditions on each side. What follows is that framework — grounded in the failure modes of synthetic data that the benchmark numbers don't show, and the cost math that most teams get wrong when they assume synthetic is always cheaper.
For the foundational case on why human feedback remains irreplaceable for alignment and safety, see why human feedback beats synthetic data for AI training. This post takes the next step: the concrete decision framework for teams that already understand the principle and need the rubric.
2. What Synthetic Data Actually Is (and Isn't)
Synthetic data is not a single thing. There are three distinct types in common use, and they have different capabilities and different failure modes. Conflating them produces confused decision-making.
(a) Model-generated text and responses as training examples
The Alpaca and Self-Instruct paradigm: use a capable model to generate instruction-response pairs at scale. GPT-4 writes the question and the answer; the smaller model trains on the result. This works for instruction diversity, general QA, and generic summarization. It does not work where the generator model's knowledge is the ceiling — if GPT-4's knowledge of a domain is shallow or systematically biased, the generated training examples are shallow and biased, and the trained model will be too.
(b) AI-generated preference labels (RLAIF)
Rather than paying humans to label which of two responses is better, you use a model to generate the preference signal. Constitutional AI's critique-revision loop uses this approach. The critical distinction: RLAIF generates labels that reflect what the labeling model prefers — not what humans actually prefer. When these are aligned, RLAIF works well. When they diverge (safety behaviors, cultural nuance, domain expertise), RLAIF amplifies the labeling model's biases rather than capturing human judgment.
(c) Augmented and paraphrased data
Using models to expand, paraphrase, or rephrase existing human-labeled data to increase volume and class distribution balance. This type carries the least risk — it is bounded by the original human-labeled distribution — but also provides the least lift. It is most useful for class imbalance problems and distribution coverage, not for injecting new domain knowledge.
What synthetic data can do: scale cheaply, cover low-risk domains, bootstrap early-stage models before a human feedback loop is in place, and generate instruction diversity at a cost that human annotation cannot match per example.
What synthetic data fundamentally cannot do: inject grounded domain expertise that the generator model does not have, catch model-specific failure modes that the generator model cannot recognize, validate real-world safety behaviors against human judgment, or provide cross-cultural nuance where the generator model's training data was not representative.
3. Where Synthetic Data Works Well
Synthetic data has proven track records in four specific contexts. If your task fits one of these, synthetic data is a legitimate first choice.
(a) Code generation bootstrapping
When you need instruction diversity fast for a code-focused model, synthetic data is highly effective. Code has verifiable correctness — you can run the generated examples and filter out the ones that don't execute. This makes synthetic code data self-correcting in a way that natural language data is not. The Phi series (Phi-1, Phi-2) demonstrated that GPT-4-generated code and reasoning examples, carefully filtered, could produce models that punched well above their parameter count on coding benchmarks.
(b) Low-stakes summarization where factual accuracy isn't critical
Generic summarization tasks — news digests, meeting note compression, product description condensation — are well-served by synthetic data where the content domain is general and errors are recoverable. If a summarization model produces a slightly inaccurate summary of a news article, the cost is low and detectable. This is categorically different from summarizing medical records or legal contracts, where errors have direct consequences.
(c) Data augmentation to balance class distributions
When you have human-labeled data but the class distribution is skewed, synthetic augmentation is a cost-effective way to balance it. Paraphrasing underrepresented class examples, generating variations on rare-event scenarios, extending low-frequency category coverage — these are legitimate uses of synthetic augmentation that improve model performance without replacing the human-labeled anchor data.
(d) Early-stage pre-RLHF training
Before you have a human feedback loop in place, synthetic data can bootstrap a model to the point where human annotation becomes productive. Gemini's training incorporated synthetic math problems to build numerical reasoning before fine-tuning with human feedback. Phi-2 used synthetic data to develop reasoning capabilities that human annotators then refined. The pattern is consistent: synthetic data for capability scaffolding, human data for alignment and quality calibration.
4. Where Synthetic Data Fails — and Why
Four failure modes. Each is structural — not a calibration problem that better prompting or more examples will fix.
(a) Reward hacking amplification
Synthetic preference labels inherit the generator model's biases. When you use RLAIF — a model labeling which response is better — the policy trained on those labels learns to optimize for what the labeling model prefers, not what humans actually want. This creates a feedback loop: the policy produces outputs the labeling model rates highly, those outputs enter the next training round, the policy drifts further toward the labeling model's preferences and further from human preferences. Anthropic's Constitutional AI paper is instructive here — CAI combines synthetic critique-revision with human preference data for safety precisely because pure synthetic alignment data is not sufficient. The synthetic component generates critique-revision pairs at scale; the human preference data anchors the alignment. Neither alone produces the result.
For the full comparison of RLHF vs. RLAIF and where AI-generated feedback diverges from human feedback, see RLHF vs. RLAIF: when AI-generated feedback isn't enough.
(b) Domain expertise gaps
GPT-4 cannot reliably annotate medical imaging, legal reasoning, or financial risk assessment at the level required for regulatory compliance or patient safety. This is not a capability limitation that the next model version will close — it is a structural consequence of the generator model's knowledge being derived from general text rather than domain practice.
Inter-annotator agreement (κ) tells the story clearly. On domain-specific tasks: crowdworkers reach κ ≈ 0.30–0.45. Domain experts reach κ ≥ 0.65. GPT-4 on specialized domain annotation tasks reaches κ ≈ 0.40–0.55 — better than crowdworkers, but still below the domain expert threshold, and with variance that depends heavily on the specific domain and task type. For medical annotation requiring clinical judgment, the gap between GPT-4 and a licensed clinician is not the gap between κ = 0.50 and κ = 0.70. It is the gap between a model that has read medical literature and a clinician who has applied it to patients.
For the structural case for domain experts over generalist annotators — including the specific κ thresholds for different domain types — see domain expert annotators vs. crowdsourcing for AI training.
(c) Evaluation blindness
Synthetic data cannot tell you whether your model works in production with real users. A model trained on GPT-4-generated QA pairs may score well on benchmarks built from GPT-4-generated evaluation sets — because both the training data and the evaluation data share the same distribution. Deploy that model to real users asking real questions in the actual distribution your product serves, and the benchmark scores become misleading. Synthetic evaluation is a closed loop: it measures how well the model has learned the synthetic distribution, not how well it performs on the human distribution you care about.
(d) Safety and alignment limitations
Pure synthetic alignment data has not produced models that satisfy human safety evaluators. Constitutional AI (Anthropic) is the most rigorous publicly documented approach to synthetic alignment — and it explicitly combines synthetic critique-revision with human preference data. The synthetic component generates critique-revision pairs that teach the model to recognize harmful content. The human preference data ensures the safety behaviors align with what humans actually consider safe, not just what the critique model flags. Models trained with synthetic alignment data only consistently fail red-teaming by human safety evaluators in ways that models trained with human preference data do not.
For a deeper treatment of synthetic vs. human data in alignment specifically, see human feedback vs. synthetic data for LLM alignment.
5. The Decision Framework — A Practical Rubric
This is the rubric. Apply it to your specific task and domain before your next planning meeting.
✓ Use synthetic data when:
- →Domain expertise is low — general QA, generic summarization, broad instruction-following diversity
- →Budget is tight and speed matters for early-stage bootstrapping
- →You need volume for pre-training diversity before a human feedback loop is in place
- →Task is low-stakes and reversible — errors are detectable and correctable post-deployment
- →You need to augment class distribution balance on an existing human-labeled dataset
✓ Use human annotation when:
- →Domain expertise is non-negotiable — medical, legal, financial, safety, and any regulated domain
- →You're doing preference optimization (RLHF) where quality of labels matters more than volume
- →You need real-world distribution coverage — actual user queries, not model-generated proxies
- →Regulatory compliance requires human-in-the-loop validation
- →You've hit a performance ceiling on synthetic data and benchmark scores are plateauing
↔ Hybrid approach (what the best models actually do)
GPT-4, Claude 3, and Llama 2 all use hybrid pipelines: synthetic data at scale for pre-training diversity and capability scaffolding, human expert data for alignment, safety, and domain quality calibration. The hybrid approach is not a compromise — it is the principled allocation of each data type to the tasks it does best. Synthetic data handles volume and distribution coverage where the generator model's knowledge is adequate. Human expert data anchors quality where it is not. The ratio varies by task; the principle is consistent.
See how Human Consensus AI's expert annotation compares
1,000 expert-annotated preference pairs with domain-matched annotators, IAA guarantees, and built-in calibration — starting at $49 for teams evaluating human annotation for the first time.
AI Training Data Starter Pack →6. The Cost Math
Synthetic data is not free, and the cost comparison is closer than most teams assume. Here is an honest accounting.
GPT-4 API costs for 10,000 preference pairs run $200–$800 depending on prompt length. At the lower end — short prompts, single-turn preference pairs — $200 buys you 10,000 labels. At the higher end — longer context, multi-turn conversations, chain-of-thought critique — $800 is realistic. Add engineering time for pipeline setup, prompt engineering, quality filtering, and format conversion, and the true cost of synthetic data is typically 2–3× the raw API cost.
Human expert annotation for 1,000 high-quality pairs from Human Consensus AI costs $49–$299 and includes domain expertise and IAA guarantees. The per-pair cost is higher than synthetic — approximately $0.05–$0.30 per annotated pair vs. $0.02–$0.08 for GPT-4-generated pairs. But 1,000 expert-annotated pairs in a domain like medical, legal, or financial AI is not the same as 1,000 GPT-4-generated pairs. The expert-annotated dataset covers real edge cases, captures domain nuance the generator model cannot represent, and comes with inter-annotator agreement scores that let you know what you are getting.
The rework cost is the number most teams miss. Discovering at evaluation time that synthetic labels were noisy or systematically biased — and then re-running the annotation pipeline with human experts — adds the full cost of the synthetic pipeline as waste. A 1,000-pair expert dataset that directly meets quality requirements is often cheaper than a 10,000-pair synthetic dataset that requires a 1,000-pair human expert correction run to be usable.
For a detailed breakdown of annotation cost structures across different task types and quality tiers, see AI annotation cost and pricing: what expert annotation actually costs.
7. The Bottom Line and How to Get Started
The framing of "synthetic vs. human" is a false binary. Most production AI teams use both. The question is not "synthetic or human?" — it is "where in my pipeline does human judgment add irreplaceable value?"
The answer is consistent across domains and model families: synthetic data is adequate for scale, diversity, and low-stakes tasks. Human expert data is required for alignment, safety, domain expertise, and the quality calibration that determines whether a model works in production or just on benchmarks.
For alignment and safety tasks, the answer is always human experts. Constitutional AI, RLHF, and every published alignment approach that has produced safe and capable models uses human preference data as the anchor. Synthetic critique-revision is a useful complement; it is not a substitute. For medical, legal, and financial AI, the answer is also always human experts — not because of principle, but because domain κ numbers make the quality gap undeniable and regulatory compliance makes it mandatory.
For teams that have been running synthetic-only pipelines and hit a performance ceiling, the path forward is not more synthetic data. It is a targeted injection of human expert annotation at the specific points where the synthetic labels are generating noisy or biased signal. A 1,000-pair expert calibration dataset, added to an existing synthetic pipeline at the preference optimization stage, consistently moves benchmark scores that had plateaued.
For the broader treatment of how human feedback interacts with synthetic data across the full LLM training pipeline, see human feedback vs. synthetic data for LLM alignment.