This landscape scope returns 229 published records dating from 2015 through the 2026 data cut-off, covering deduplication methods like minhash and locality-sensitive hashing together with curation techniques such as benchmark-contamination and toxicity filtering. That figure counts records, and the same 229 records form the denominator for every share quoted on this page. It is not a global count of every deduplication patent ever filed — it reflects the specific search terms and date range used to build this dataset.

The ranking covers 80 companies across the full 229-record scope, with the leading assignee holding 16 records and the top five together holding 63 records, 27.5% of the total. That leaves a substantial long tail of companies with only one or two filings each, so no single company controls the fundamental mechanisms. Reviewing the leader’s specific claim language is more useful than assuming the field is locked up by a handful of firms.

Filing peaked at 21 records in 2019 and, over the most recently comparable three-year window, moved from 16 records in 2021 to 12 in 2024 — a 25% decline. However, publication typically lags filing by around 18 months, so 2025 and 2026 figures are still incomplete and should not be read as a continuation of that decline, especially given the recent growth in AI-training data curation work. A fair read is that the field cooled after 2019 but the newest activity has not fully surfaced yet.