Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errorsOnly 34% of enterprise leaders surveyed in early 2026 said synthetic data provides referential integrity. That figure measures confidence, not a laboratory failure rate: it is the share of 518 global leaders who gave that answer. The same survey found that 36% trusted synthetic data for data realism, even as 56% considered it the most purpose-fit approach for protecting AI/ML workflows.
This gap between strategic appeal and confidence helps explain why synthetic data is used widely in some teams but remains absent from many AI/ML and software-delivery pipelines.
As an Amazon Associate I earn from qualifying purchases.
What the 34% figure actually measures
Perforce Delphix’s 2026 survey covered 518 enterprise leaders worldwide, including people working in cybersecurity, compliance, software engineering and software testing. Respondents were asked whether different data-protection approaches provide specific capabilities. Only 34% said synthetic data provides referential integrity, and 36% said it provides data realism.
Those are perception results, not an independently verified benchmark, defect rate or claim that 66% of synthetic-data records are invalid. They show that a large majority of surveyed leaders do not yet trust the synthetic-data approaches available to them to preserve relationships reliably.
#1 Best Overall
Why referential integrity matters beyond a database schema
Referential integrity means that identifiers and relationships remain valid: a customer referenced by an order still exists, a transaction still points to the correct account, and the same entity can be joined consistently across tables, applications and databases. Enterprise analytics, integration tests and AI pipelines depend on those links to assemble complete populations.
When relationships break, a pipeline may still run and produce apparently plausible output. The result can be incomplete populations, drifting metrics and silent differences between systems. Perforce Delphix’s 2026 State of AI and Data Privacy Report warns that flawed or incomplete data can increase bias, degrade model performance and make failures harder to diagnose. Broken integrity also erodes confidence in analytics and AI, pushing teams back toward manual checks or narrower, less representative test data.
The adoption contradiction
Synthetic data has a strong strategic rationale for privacy-sensitive AI/ML work, but confidence and actual deployment lag behind that interest.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Rank #2
| 2026 survey result | Share | What it indicates |
|---|---|---|
| Leaders saying synthetic data provides referential integrity | 34% | Confidence in preserving relationships is limited |
| Leaders saying synthetic data provides data realism | 36% | Realistic business behavior is not widely trusted |
| Leaders ranking synthetic data the most purpose-fit protection for AI/ML workflows | 56% | Strategic fit is higher than confidence in quality |
| Organizations not using synthetic data in AI/ML workflows | 51% | Interest has not translated into routine use |
| Software leaders not using it in development and testing | 66% | Delivery teams remain particularly cautious |
The contrast is not necessarily irrational. A method can be attractive for reducing exposure to production data while still requiring proof that keys, joins, distributions and edge cases survive transformation. Teams often postpone adoption when the cost of diagnosing a broken relationship is greater than the privacy benefit of switching methods.
Where organizations are using synthetic data today
Perforce’s 2025 State of Synthetic Data Report found that 63% of organizations used synthetic data in some capacity. Among those users, the leading workloads were analytics and integration testing.
| Workload | Share of synthetic-data users |
|---|---|
| Analytics | 79% |
| Integration testing | 74% |
| Software development | 63% |
| Software testing | 62% |
Integration testing has a particularly demanding requirement: protection rules must remain consistent across systems, and the resulting data must mirror production relationships and complexity. A dataset that looks realistic in one application can still fail when an order, identity, payment or entitlement must be correlated with records in another database.
How synthetic data can preserve relationships—and where it can fail
Synthetic generation can preserve relationships only when the generation process treats them as explicit constraints rather than creating each table independently. A dependable design needs a shared entity and key strategy, cross-system mapping rules, foreign-key validation and business-rule checks for cases such as lifecycle states, dates and cardinality. Those controls must be tested against the actual workload; a visually realistic row is not evidence that its joins are correct.
Common failure patterns
- Independent generation: related tables receive different identifiers, so joins silently lose records.
- Partial population: a generated subset omits parent or history records needed to represent a complete process.
- Rule drift: one system changes a status or code list while another keeps an older version.
- Distribution loss: generated values preserve formats but not correlations, seasonality or rare business cases.
- Environment mismatch: data that works in a single database does not reproduce dependencies across services.
How the alternatives scored on integrity
The 2026 comparison asked leaders which approaches provide referential integrity within and across databases. The percentages below are perceptions from that survey, not universal technical guarantees.
| Approach | Leaders saying it provides referential integrity | Important qualification |
|---|---|---|
| Static masking | 63% | Highest perceived integrity in this comparison; transformed copies still require validation |
| Dynamic masking | 36% | Perception was close to synthetic data’s 34% |
| Synthetic data | 34% | Perception measure, not a failure rate |
| Data subsetting | 18% | Smaller datasets can omit dependencies unless selection rules include them |
| Tokenization | 17% | Integrity depends on consistent token mapping across every system |
The choice should follow the workload rather than the ranking alone. Evaluate referential integrity alongside realism, statistical fidelity, privacy and re-identification risk, scale and delivery time, operational effort, and the maintenance required as schemas and business rules change. No result in this survey establishes that one approach solves every cross-database workload.
Rank #4
A practical acceptance test before adopting synthetic data
- Inventory relationships. Document primary and foreign keys, shared identifiers, cross-database joins, event order and required parent-child cardinalities.
- Define workload-specific fixtures. Include normal transactions, deletes, corrections, retries, late-arriving records, rare states and intentionally invalid inputs that the system must reject.
- Set measurable checks. Test orphan rates, join completeness, row-count tolerances, duplicate keys, code-list validity, time relationships and aggregate totals.
- Validate statistical behavior. Compare distributions and correlations that affect model training, reporting or test decisions; format similarity alone is insufficient.
- Test privacy separately. Confirm that the transformation reduces disclosure and re-identification risk for the intended environment and access model.
- Run at production-like scale. Measure generation, refresh and delivery against the team’s service-level requirements, including failure recovery.
- Re-test after change. Treat schema migrations, new services and changed business rules as triggers for renewed integrity and realism tests.
Governance is part of the trust problem
A 2025 UK Administrative Data Research UK report, based on a literature review, surveys, case studies and a focus group with Trusted Research Environment professionals, identified legal ambiguity, limited resources, inconsistent quality-assurance practices and misconceptions as barriers to synthetic-data adoption in research settings. These issues apply to enterprise decisions as well: confidence depends on documented controls, ownership and repeatable validation, not only on the generator’s output.
Teams should record which relationships were tested, which workloads passed, what privacy assumptions apply and how failures are handled. That evidence gives security, compliance, engineering and model-risk stakeholders a common basis for approving a use case or choosing a different protection method.
What the 34% result means for decision-makers
Synthetic data is neither broadly untrustworthy nor automatically safe for joins. It is already used by a majority of organizations in at least one capacity, especially analytics and integration testing, yet many leaders still lack proof that it preserves relationships across their own systems. The sensible decision is workload-specific: require demonstrable keys, joins, business rules, realism, privacy, scale and recovery behavior before replacing production-derived data, and keep an alternative such as masking when those tests do not pass.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
