Analysis of 14M tweets from May 2016 and March 2017 found bots accounted for 34% of shares of articles from low-credibility sources but were only 6% of accounts
Daniel Funke / Poynter : Tweets: @brendannyhan See also Mediagazer Tweets: Brendan Nyhan / @brendannyhan : The latest on bot prevalence and effects: -Bots increase exposure to negative & inflammatory content in online social systems (@PNASNews) http://www.pnas.org/... -Bots spread a lot of fakery during the 2016 election. But they can also debunk it. https://www.poynter.org/... (@poynter) See also Mediagazer
Context & Ripple Effects
This analysis lands mid-arc in a running argument about who actually spreads junk news. Earlier work cut both ways: Oxford researchers found Twitter bots pushed misinfo at higher rates in battleground states around Election Day, while a later study of 4.5M+ tweets concluded humans, not bots, were primarily responsible for inaccurate news spreading faster than true stories.
First-order effects
- The 34%-of-shares vs 6%-of-accounts ratio gives Poynter, Nyhan, and fellow researchers a precise reconciliation: bots are a tiny minority of accounts but roughly six times overrepresented in low-credibility article sharing, so both camps' findings can hold at once — humans supply the volume, bots concentrate on junk.
Second-order effects
- Platforms now face pressure to police amplification rather than raw bot counts, since Pew's earlier finding that 500 suspected bot accounts drove 22% of shared news links shows a handful of accounts can dominate distribution metrics that newsrooms and advertisers treat as audience signal.
Third-order effects
- If the pattern holds, the policy debate shifts from 'how many bots exist' to 'what share of reach they command,' pushing regulators and platforms toward amplification-weighted transparency rules — with public opinion already primed, as Pew found about 66% of US adults aware of social media bots and most assuming malicious use.
The trend: Misinformation research is converging on bots as a small-account, high-amplification layer atop human sharing, moving the field from measuring bot prevalence to measuring bot leverage.