This one is gloriously semantic. The critical bit is adding the qualifying words "abnormal" and "risk-adjusted".
Imagine for example's sake a market stylized by two groups of stocks: debt-heavy telecoms companies, and cash-rich tech companies. A change in credit spreads will obviously work to the relative benefit of one group, and to the relative cost of the other. Which is saying nothing more than different stocks have different credit betas.
For the EMH to hold, you have to believe that all available information is reflected in the current valuation of credit spreads; and thus the future path of credit spreads is unpredictable.
Future news and the surprise this represents will cause credit spreads to move. The effects this will have on different stocks is (relatively) predictable. But putting a predictable multiplier on a random input still gives you a random output. Plus the residuals of my companies' "alpha" excluding these effects, will still (in theory) be random noise. If return = beta * market + error, then the EMH holds if market and error are both random.
When I go and buy dinner with my returns, neither I nor the supermarket care much whether these were generated by alpha, this beta, or that one. However, people who write textbooks seem to get hot under the collar about such things ;-)