We empirically investigate the participation and transactional liquidity provided by algorithmic vs. human traders during “abnormally” stressful periods, relative to what they do in “normal” periods, and the resultant implications for the quality and fragility of markets. We find strong evidence that, in periods of abnormal stress, algorithmic traders significantly reduce their participation and liquidity provision in trades; significantly reduce the extent to which they post new liquidity-supplying limit orders; significantly reduce the aggressiveness of these limit orders, and sharply increase the price at which they are willing to supply liquidity. We define abnormal stress based on persistently extreme levels of volatility, order-imbalances, and bid-offer spreads; and measures that proxy for “ambiguity” or complexity. This significantly greater withdrawal of algorithmic (relative to human) traders is directly associated with the disappearance of information advantages of algorithmic traders. We find that this has a significant propensity to generate feedback loops, and induce “contagion” through withdrawals in liquidity provision in related stocks, potentially making markets more “fragile”. Our results suggest that, in contrast to manual traders who adapt in (higher latency) real time, algorithmic trade execution appears less conducive to low impact adjustment of ambiguous information asymmetries or flows. Overall, our results reinforce regulatory concerns about the potential for systemic fragility in markets dominated by machine-based liquidity provision.